(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control... · Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (2024)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (1)

Slotine • Li APPLIED NONLINEAR CONTROL

! i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (2)

APPLIEDNONLINEAR

CONTROL

Jean-Jacques E SlotineWeiping Li

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (3)

AppliedNonlinearControl

JEAN-JACQUES E. SLOTINEMassachusetts Institute of Technology

WEIPING LIMassachusetts Institute of Technology'

Prentice HallEnglewood Cliffs, New Jersey 07632

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (4)

Library of Congress Cataloging-in-Publication Data

Slotine, J.-J. E. (Jean-Jacques E.)Applied nonlinear control / Jean-Jacques E. Slotine, Weiping Li

p. cm.

Includes bibliographical references.

ISBN 0-13-040890-5

1, Nonlinear control theory. I. Li, Weiping. II. Title.

QA402.35.S56 1991 90-33365629.8'312-dc20 C1P

Editorial/production supervision andinterior design: JENNIFER WENZEL

Cover design: KAREN STEPHENSManufacturing Buyer: LORI BULWIN

= ^ = © 1991 by Prentice-Hall, Inc.^=&= A Division of Simon & Schuster

T k Englewood Cliffs, New Jersey 07632

All rights reserved. No part of this book may bereproduced, in any form or by any means,without permission in writing from the publisher.

Printed in the United States of America

20 19 18 17 16 15 14 13 12 1]

ISBN D-13-DHDfiTa-S

Prentice-Hall International (UK) Limited, LondonPrentice-Hall of Australia Pty. Limited, SydneyPrentice-Hall Canada Inc., TorontoPrentice-Hail Hispanoamericana, S.A., MexicoPrentice-Hall of India Private Limited, New DelhiPrentice-Hall of Japan, Inc., TokyoSimon & Schuster Asia Pte. Ltd., SingaporeEditora Prentice-Hall do Brasil, Ltda., Rio de Janeiro

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (5)

To Our Parents

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (6)

Contents

Preface xi

1. Introduction 1

1.1 Why Nonlinear Control ? 1

1.2 Nonlinear System Behavior 4

1.3 An Overview of the Book 12

1.4 Notes and References 13

Part I: Nonlinear Systems Analysis 14Introduction to Part I 14

2. Phase Plane Analysis 17

2.1 Concepts of Phase Plane Analysis 182.1.1 Phase Portraits 182.1.2 Singular Points 202.1.3 Symmetry in Phase Plane Portraits 22

2.2 Constructing Phase Portraits 23

2.3 Determining Time from Phase Portraits 29

2.4 Phase Plane Analysis of Linear Systems 30

2.5 Phase Plane Analysis of Nonlinear Systems 32

2.6 Existence of Limit Cycles 36

2.7 Summary 38

2.8 Notes and References 38

2.9 Exercises 38

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (7)

VI11

3. Fundamentals of Lyapunov Theory 40

3.1 Nonlinear Systems and Equilibrium Points 41

3.2 Concepts of Stability 47

3.3 Linearization and Local Stability 53

3.4 Lyapunov's Direct Method 573.4.1 Positive Definite Functions and Lyapunov Functions 583.4.2 Equilibrium Point Theorems 613.4.3 Invariant Set Theorems 68

3.5 System Analysis Based on Lyapunov's Direct Method 763.5.1 Lyapunov Analysis of Linear Time-Invariant Systems 773.5.2 Krasovskii's Method 833.5.3 The Variable Gradient Method 863.5.4 Physically Motivated Lyapunov Functions 883.5.5 Performance Analysis 91

3.6 Control Design Based on Lyapunov's Direct Method 94

3.7 Summary 95

3.8 Notes and References 96

3.9 Exercises 97

4. Advanced Stability Theory 100

4.1 Concepts of Stability for Non-Autonomous Systems 101

4.2 Lyapunov Analysis of Non-Autonomous Systems 105

4.2.1 Lyapunov's Direct Method for Non-Autonomous Systems 1054.2.2 Lyapunov Analysis of Linear Time-Varying Systems 114

4.2.3 The Linearization Method for Non-Autonomous Systems 116

4.3 * Instability Theorems 117

4.4 * Existence of Lyapunov Functions 1204.5 Lyapunov-Like Analysis Using Barbalat's Lemma 122

4.5.1 Asymptotic Properties of Functions and Their Derivatives 1224.5.2 Barbalat's Lemma 123

4.6 Positive Linear Systems 1264.6.1 PR and SPR Transfer Functions 1264.6.2 The Kalman-Yakubovich Lemma 1304.6.3 Positive Real Transfer Matrices 131

4.7 The Passivity Formalism 1324.7.1 Block Combinations 1324.7.2 Passivity in Linear Systems 137

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (8)

IX

4.8 * Absolute Stability 142

4.9 * Establishing Boundedness of Signals 147

4.10 * Existence and Unicity of Solutions 151

4.11 Summary 153

4.12 Notes and References 153

4.13 Exercises 154

5. Describing Function Analysis 157

5.1 Describing Function Fundamentals 1585.1.1 An Example of Describing Function Analysis 1585.1.2 Applications Domain 1625.1.3 Basic Assumptions 1645.1.4 Basic Definitions 1655.1.5 Computing Describing Functions 167

5.2 Common Nonlinearities In Control Systems 169

5.3 Describing Functions of Common Nonlinearities 172

5.4 Describing Function Analysis of Nonlinear Systems 1795.4.1 The Nyquist Criterion and Its Extension 1805.4.2 Existence of Limit Cycles 1825.4.3 Stability of Limit Cycles 1845.4.4 Reliability of Describing Function Analysis 186

5.5 Summary 187

5.6 Notes and References 188

5.7 Exercises 188

Part II: Nonlinear Control Systems Design 191Introduction to Part II 191

6. Feedback Linearization 207

6.1 Intuitive Concepts 2086.1.1 Feedback Linearization And The Canonical Form 2086.1.2 Input-State Linearization 2136.1.3 Input-Output Linearization 216

6.2 Mathematical Tools 229

6.3 Input-State Linearization of SISO Systems 236

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (9)

6.4 Input-Output Linearization of SISO Systems 246

6.5 * Multi-Input Systems 266

6.6 Summary 270

6.7 Notes and References 271

6.8 Exercises 271

7. Sliding Control 276

7.1 Sliding Surfaces 2777.1.1 A Notational Simplification 2787.1.2 * Filippov's Construction of the Equivalent Dynamics 2837.1.3 Perfect Performance - At a Price 2857.1.4 Direct Implementations of Switching Control Laws 289

7.2 Continuous Approximations of Switching Control Laws 290

7.3 The Modeling/Performance Trade-Offs 301

7.4 * Multi-Input Systems 303

7.5 Summary 306

7.6 Notes and References 307

7.7 Exercises 307

8. Adaptive Control 311

8.1 Basic Concepts in Adaptive Control 3128.1.1 Why Adaptive Control ? 3128.1.2 What Is Adaptive Control ? 3158.1.3 How To Design Adaptive Controllers ? 323

8.2 Adaptive Control of First-Order Systems 326

8.3 Adaptive Control of Linear Systems With Full State Feedback 335

8.4 Adaptive Control of Linear Systems With Output Feedback 339

8.4.1 Linear Systems With Relative Degree One 340

8.4.2 Linear Systems With Higher Relative Degree 346

8.5 Adaptive Control of Nonlinear Systems 350

8.6 Robustness of Adaptive Control Systems 3538.7 * On-Line Parameter Estimation 358

8.7.1 Linear Parametrization Model 3598.7.2 Prediction-Error-Based Estimation Methods 3648.7.3 The Gradient Estimator 3648.7.4 The Standard Least-Squares Estimator 370

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (10)

8.7.5 Least-Squares With Exponential Forgetting 3748.7.6 Bounded-Gain Forgetting 3768.7.7 Concluding Remarks and Implementation Issues 381

1.8 Composite Adaptation 382

1.9 Summary 388

1.10 Notes and References 389

1.11 Exercises 389

9. Control of Multi-Input Physical Systems 392

9.1 Robotics as a Prototype 3939.1.1 Position Control 3949.1.2 Trajectory Control 397

9.2 Adaptive Robot Trajectory Control 403

9.2.1 The Basic Algorithm 404

9.2.2 * Composite Adaptive Trajectory Control 411

9.3 Putting Physics in Control 416

9.3.1 High-Frequency Unmodeled Dynamics 4169.3.2 Conservative and Dissipative Dynamics 4189.3.3 Robotics as a Metaphor 419

9.4 Spacecraft Control 422

9.4.1 The Spacecraft Model 422

9.4.2 Attitude Control 425

9.5 Summary 432

9.6 Notes and References 433

9.7 Exercises 433

BIBLIOGRAPHY 437

INDEX 459

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (11)

Preface

In recent years, the availability of powerful low-cost microprocessors has spurredgreat advances in the theory and applications of nonlinear control. In terms of theory,major strides have been made in the areas of feedback linearization, sliding control,and nonlinear adaptation techniques. In terms of applications, many practicalnonlinear control systems have been developed, ranging from digital "fly-by-wire"flight control systems for aircraft, to "drive-by-wire" automobiles, to advanced roboticand space systems. As a result, the subject of nonlinear control is occupying anincreasingly important place in automatic control engineering, and has become anecessary part of the fundamental background of control engineers.

This book, based on a course developed at MIT, is intended as a textbook forsenior and graduate students, and as a self-study book for practicing engineers. Itsobjective is to present the fundamental results of modern nonlinear control whilekeeping the mathematical complexity to a minimum, and to demonstrate their use andimplications in the design of practical nonlinear control systems. Although a majormotivation of this book is to detail the many recent developments in nonlinear control,classical techniques such as phase plane analysis and the describing function methodare also treated, because of their continued practical importance.

In order to achieve our fundamental objective, we have tried to bring thefollowing features to this book:

• Readability: Particular attention is paid to the readability of the book bycarefully organizing the concepts, intuitively interpreting the major results, andselectively using the mathematical tools. The readers are only assumed to have hadone introductory control course. No mathematical background beyond ordinarydifferential equations and elementary matrix algebra is required. For each newresult, interpretation is emphasized rather than mathematics. For each major result,we try to ask and answer the following key questions: What does the resultintuitively and physically mean? How can it be applied to practical problems?What is its relationship to other theorems? All major concepts and results aredemonstrated by examples. We believe that learning and generalization fromexamples are crucial for proficiency in applying any theoretical result.

• Practicality: The choice and emphasis of materials is guided by the basic

xiii

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (12)

XIV

objective of making an engineer or student capable of dealing with practical controlproblems in industry. Some results of mostly theoretical interest are not included.The selected materials, in one way or another, are intended to allow readers to gaininsights into the solution of real problems.

• Comprehensiveness: The book contains both classical materials, such asLyapunov analysis and describing function techniques, and more modern topicssuch as feedback linearization, adaptive control, and sliding control. To facilitatedigestion, asterisks are used to indicate sections which, given their relativecomplexity, can be safely skipped in a first reading.

• Currentness: In the past few years, a number of major results have beenobtained in nonlinear control, particularly in nonlinear control system design and inrobotics. It is one of the objectives of this book to present these new and importantdevelopments, and their implications, in a clear, easily understandable fashion.The book can thus be used as a reference and a guide to the active literature inthese fields.

The book is divided into two major parts. Chapters 2-5 present the majoranalytical tools that can be used to study a nonlinear system, while chapters 6-9 treatthe major nonlinear controller design techniques. Each chapter is supplied withexercises, allowing the reader to further explore specific aspects of the materialdiscussed. A detailed index and a bibliography are provided at the end of the book.

The material included exceeds what can be taught in one semester or self-learned in a short period. The book can be studied in many ways, according to theparticular interests of the reader or the instructor. We recommend that a first readinginclude a detailed study of chapter 3 (basic Lyapunov theory), sections 4.5-4.7(Barbalat's lemma and passivity tools), section 6.1 and parts of sections 6.2-6.4(feedback linearization), chapter 7 (sliding control), sections 8.1-8.3 and 8.5 (adaptivecontrol of linear and nonlinear systems), and chapter 9 (control of multi-input physicalsystems). Conversely, sections denoted with an asterisk can be skipped in a firstreading.

Many colleagues, students, and friends greatly contributed to this book throughstimulating discussions and judicious suggestions. Karl Hedrick provided us withcontinued enthusiasm and encouragement, and with many valuable comments andsuggestions. Discussions with Karl Astrdm and Semyon Meerkov helped us betterdefine the tone of the book and its mathematical level. Harry Asada, Jo Bentsman,Marika DiBenedetto, Olav Egeland, Neville Hogan, Marija Ilic, Lars Nielsen, KenSalisbury, Sajhendra Singh, Mark Spong, David Wormley, and Dana Yoergerprovided many useful suggestions and much moral support. Barbara Hove created

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (13)

XV

most of the nicer drawings in the book; Giinter Niemeyer's expertise and energy wasinvaluable in setting up the computing and word processing environments; Hyun Yanggreatly helped with the computer simulations; all three provided us with extensivetechnical and editorial comments. The book also greatly benefited from the interestand enthusiasm of many students who took the course at MIT.

Partial summer support for the first author towards the development of the bookwas provided by Gordon Funds. Finally, the energy and professionalism of Tim Bozikand Jennifer Wenzel at Prentice-Hall were very effective and highly appreciated.

Jean-Jacques E. SlotineWeiping Li

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (14)

AppliedNonlinearControl

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (15)

Chapter 1Introduction

The subject of nonlinear control deals with the analysis and the design of nonlinearcontrol systems, i.e., of control systems containing at least one nonlinear component.In the analysis, a nonlinear closed-loop system is assumed to have been designed, andwe wish to determine the characteristics of the system's behavior. In the design, weare given a nonlinear plant to be controlled and some specifications of closed-loopsystem behavior, and our task is to construct a controller so that the closed loopsystem meets the desired characteristics. In practice, of course, the issues of designand analysis are intertwined, because the design of a nonlinear control system usuallyinvolves an iterative process of analysis and design.

This introductory chapter provides the background for the specific analysis anddesign methods to be discussed in the later chapters. Section 1.1 explains themotivations for embarking on a study of nonlinear control. The unique and richbehaviors exhibited by nonlinear systems are discussed in section 1.2. Finally, section1.3 gives an overview of the organization of the book.

1.1 Why Nonlinear Control ?

Linear control is a mature subject with a variety of powerful methods and a longhistory of successful industrial applications. Thus, it is natural for one to wonder whyso many researchers and designers, from such broad areas as aircraft and spacecraftcontrol, robotics, process control, and biomedical engineering, have recently showed

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (16)

2 Introduction Chap. 1

an active interest in the development and applications of nonlinear controlmethodologies. Many reasons can be cited for this interest:

• Improvement of existing control systems: Linear control methods rely onthe key assumption of small range operation for the linear model to be valid. Whenthe required operation range is large, a linear controller is likely to perform verypoorly or to be unstable, because the nonlinearities in the system cannot be properlycompensated for. Nonlinear controllers, on the other hand, may handle thenonlinearities in large range operation directly. This point is easily demonstrated inrobot motion control problems. When a linear controller is used to control robotmotion, it neglects the nonlinear forces associated with the motion of the robot links.The controller's accuracy thus quickly degrades as the speed of motion increases,because many of the dynamic forces involved, such as Coriolis and centripetal forces,vary as the square of the speed. Therefore, in order to achieve a pre-specifiedaccuracy in robot tasks such as pick-and-place, arc welding and laser cutting, thespeed of robot motion, and thus productivity, has to be kept low. On the other hand, aconceptually simple nonlinear controller, commonly called computed torquecontroller, can fully compensate the nonlinear forces in the robot motion and lead tohigh accuracy control for a very large range of robot speeds and a large workspace.

• Analysis of hard nonlinearities: Another assumption of linear control is thatthe system model is indeed linearizable. However, in control systems there are manynonlinearities whose discontinuous nature does not allow linear approximation. Theseso-called "hard nonlinearities" include Coulomb friction, saturation, dead-zones,backlash, and hysteresis, and are often found in control engineering. Their effectscannot be derived from linear methods, and nonlinear analysis techniques must bedeveloped to predict a system's performance in the presence of these inherentnonlinearities. Because such nonlinearities frequently cause undesirable behavior ofthe control systems, such as instabilities or spurious limit cycles, their effects must bepredicted and properly compensated for.

• Dealing with model uncertainties: In designing linear controllers, it isusually necessary to assume that the parameters of the system model are reasonablywell known. However, many control problems involve uncertainties in the modelparameters. This may be due to a slow time variation of the parameters (e.g., ofambient air pressure during an aircraft flight), or to an abrupt change in parameters(e.g,, in the inertial parameters of a robot when a new object is grasped). A linearcontroller based on inaccurate or obsolete values of the model parameters may exhibitsignificant performance degradation or even instability. Nonlinearities can beintentionally introduced into the controller part of a control system so that model

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (17)

Sect. 1.1 Why Nonlinear Control? 3

uncertainties can be tolerated. Two classes of nonlinear controllers for this purposeare robust controllers and adaptive controllers.

• Design Simplicity: Good nonlinear control designs may be simpler and moreintuitive than their linear counterparts. This a priori paradoxical result comes from thefact that nonlinear controller designs are often deeply rooted in the physics of theplants. To take a very simple example, consider a swinging pendulum attached to ahinge, in the vertical plane. Starting from some arbitrary initial angle, the pendulumwill oscillate and progressively stop along the vertical. Although the pendulum'sbehavior could be analyzed close to equilibrium by linearizing the system, physicallyits stability has very little to do with the eigenvalues of some linearized system matrix:it comes from the fact that the total mechanical energy of the system is progressivelydissipated by various friction forces (e.g., at the hinge), so that the pendulum comes torest at a position of minimal energy.

There may be other related or unrelated reasons to use nonlinear controltechniques, such as cost and performance optimality. In industrial settings, ad-hocextensions of linear techniques to control advanced machines with significantnonlinearities may result in unduly costly and lengthy development periods, where thecontrol code comes with little stability or performance guarantees and is extremelyhard to transport to similar but different applications. Linear control may require highquality actuators and sensors to produce linear behavior in the specified operationrange, while nonlinear control may permit the use of less expensive components withnonlinear characteristics. As for performance optimality, we can cite bang-bang typecontrollers, which can produce fast response, but are inherently nonlinear.

Thus, the subject of nonlinear control is an important area of automatic control.Learning basic techniques of nonlinear control analysis and design can significantlyenhance the ability of a control engineer to deal with practical control problemseffectively. It also provides a sharper understanding of the real world, which isinherently nonlinear. In the past, the application of nonlinear control methods hadbeen limited by the computational difficulty associated with nonlinear control designand analysis. In recent years, however, advances in computer technology have greatlyrelieved this problem. Therefore, there is currently considerable enthusiasm for theresearch and application of nonlinear control methods. The topic of nonlinear controldesign for large range operation has attracted particular attention because, on the onehand, the advent of powerful microprocessors has made the implementation ofnonlinear controllers a relatively simple matter, and, on the other hand, moderntechnology, such as high-speed high-accuracy robots or high-performance aircrafts, isdemanding control systems with much more stringent design specifications.Nonlinear control occupies an increasingly conspicuous position in control

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (18)

4 Introduction Chap. 1

engineering, as reflected by the ever-increasing number of papers and reports onnonlinear control research and applications.

1.2 Nonlinear System Behavior

Physical systems are inherently nonlinear. Thus, all control systems are nonlinear to acertain extent. Nonlinear control systems can be described by nonlinear differentialequations. However, if the operating range of a control system is small, and if theinvolved nonlinearities are smooth, then the control system may be reasonablyapproximated by a linearized system, whose dynamics is described by a set of lineardifferential equations.

NONLINEARITIES

Nonlinearities can be classified as inherent (natural) and intentional (artificial).Inherent nonlinearities are those which naturally come with the system's hardware andmotion. Examples of inherent nonlinearities include centripetal forces in rotationalmotion, and Coulomb friction between contacting surfaces. Usually, suchnonlinearities have undesirable effects, and control systems have to properlycompensate for them. Intentional nonlinearities, on the other hand, are artificiallyintroduced by the designer. Nonlinear control laws, such as adaptive control laws andbang-bang optimal control laws, are typical examples of intentional nonlinearities.

Nonlinearities can also be classified in terms of their mathematical properties,as continuous and discontinuous. Because discontinuous nonlinearities cannot belocally approximated by linear functions, they are also called "hard" nonlinearities.Hard nonlinearities (such as, e.g., backlash, hysteresis, or stiction) are commonlyfound in control systems, both in small range operation and large range operation.Whether a system in small range operation should be regarded as nonlinear or lineardepends on the magnitude of the hard nonlinearities and on the extent of their effectson the system performance. A detailed discussion of hard nonlinearities is provided insection 5.2.

LINEAR SYSTEMS

Linear control theory has been predominantly concerned with the study of linear time-invariant (LTI) control systems, of the form

x = Ax (1.1)

with x being a vector of states and A being the system matrix. LTI systems have quitesimple properties, such as

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (19)

Sect. 1.2 Nonlinear System Behavior 5

• a linear system has a unique equilibrium point if A is nonsingular;

• the equilibrium point is stable if all eigenvalues of A have negative real

parts, regardless of initial conditions;

• the transient response of a linear system is composed of the natural modes

of the system, and the general solution can be solved analytically;

• in the presence of an external input u(t), i.e., with

x=Ax+Bu , (1.2)

the system response has a number of interesting properties. First, it satisfies

the principle of superposition. Second, the asymptotic stability of the system

(1.1) implies bounded-input bounded-output stability in the presence of u.

Third, a sinusoidal input leads to a sinusoidal output of the same frequency.

AN EXAMPLE OF NONLINEAR SYSTEM BEHAVIOR

The behavior of nonlinear systems, however, is much more complex. Due to the lack

of linearity and of the associated superposition property, nonlinear systems respond to

external inputs quite differently from linear systems, as the following example

illustrates.

Example 1.1: A simplified model of the motion of an underwater vehicle can be written

v + |v|v = « (1.3)

where v is the vehicle velocity and u is the control input (the thrust provided by a propeller). The

nonlinearity |v|v corresponds to a typical "square-law" drag.

Assume that we apply a unit step input in thrust u, followed 5 seconds later by a negative unit

step input. The system response is plotted in Figure 1.1. We see that the system settles much

faster in response to the positive unit step than it does in response to the subsequent negative unit

step. Intuitively, this can be interpreted as reflecting the fact that the "apparent damping"

coefficient |v| is larger at high speeds than at low speeds.

Assume now that we repeat the same experiment but with larger steps, of amplitude 10.

Predictably, the difference between the settling times in response to the positive and negative

steps is even more marked (Figure 1.2). Furthermore, the settling speed vs in response to the first

step is not 10 times that obtained in response to the first unit step in the first experiment, as it

would be in a linear system. This can again be understood intuitively, by writing that

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (20)

Introduction

3 1.0.

Chap. 1

•5 0.8

0.6

0.4

0.2

0.0

10.0

5 10time(sec)

Figure 1.1 : Response of system (1.3) to unit steps

> 10.0 r

10

8.0

6.0

4.0

2.0

00

-

-

t

10time(sec)

ity8 8.0

6.0

4.0

2.0

0.010

15 20time(sec)

15 20time(sec)

Figure 1.2 : Response of system (1.3) to steps of amplitude 10

« = 1 = > 0 + | v s | v J = l = > V j = l

u = 1 0 => 0 + l v J v ^ l O => vi

Carefully understanding and effectively controlling this nonlinear behavior is particularly

important if the vehicle is to move in a large dynamic range and change speeds continually, as is

typical of industrial remotely-operated underwater vehicles (R.O.V.'s). D

SOME COMMON NONLINEAR SYSTEM BEHAVIORS

Let us now discuss some common nonlinear system properties, so as to familiarize

ourselves with the complex behavior of nonlinear systems and provide a useful

background for our study in the rest of the book.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (21)

Sect. 1.2

Multiple Equilibrium Points

Nonlinear System Behavior

Nonlinear systems frequently have more than one equilibrium point (an equilibriumpoint is a point where the system can stay forever without moving, as we shallformalize later). This can be seen by the following simple example.

Example 1.2: A first-order system

Consider the first order system

with initial condition x(0) = xg. Its linearization is

x = -x

(1.4)

(1.5)

The solution of this linear equation is x(t) =xoe~'. It is plotted in Figure 1.3(a) for various initial

conditions. The linearized system clearly has a unique equilibrium point at x = 0.

By contrast, integrating equation dx/(-x + x2) - dt, the actual response of the nonlinear

dynamics (1.4) can be found to be

x(t) = -

This response is plotted in Figure 1.3(b) for various initial conditions. The system has two

equilibrium points, x - 0 and x = 1 , and its qualitative behavior strongly depends on its initial

condition. CD

x(t) \x(t)

(a) (b)

Figure 1.3 : Responses of the linearized system (a) and the nonlinear system (b)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (22)

8 Introduction Chap. 1

The issue of motion stability can also be discussed with the aid of the aboveexample. For the linearized system, stability is seen by noting that for any initialcondition, the motion always converges to the equilibrium point x = 0. However,consider now the actual nonlinear system. While motions starting with xo < 1 willindeed converge to the equilibrium point x = 0, those starting with xo> I will go toinfinity (actually in finite time, a phenomenon known as finite escape time). Thismeans that the stability of nonlinear systems may depend on initial conditions.

In the presence of a bounded external input, stability may also be dependent onthe input value. This input dependence is highlighted by the so-called bilinear system

X = XU

If the input u is chosen to be - 1, then the state x converges to 0. If « = 1, then | x \tends to infinity.

Limit Cycles

Nonlinear systems can display oscillations of fixed amplitude and fixed period withoutexternal excitation. These oscillations are called limit cycles, or self-excitedoscillations. This important phenomenon can be simply illustrated by a famousoscillator dynamics, first studied in the 1920's by the Dutch electrical engineerBalthasar Van der Pol.

Example 1.3: Van der Pol Equation

The second-order nonlinear differential equation

mx + 2c(x2-l)x + kx = 0 (1.6)

where m, c and k are positive constants, is the famous Van der Pol equation. It can be regarded as

describing a mass-spring-damper system with a position-dependent damping coefficient

2c(x2 - 1) (or, equivalently, an RLC electrical circuit with a nonlinear resistor). For large values

of jr, the damping coefficient is positive and the damper removes energy from the system. This

implies that the system motion has a convergent tendency. However, for small values of x, the

damping coefficient is negative and the damper adds energy into the system. This suggests that

the system motion has a divergent tendency. Therefore, because the nonlinear damping varies

with x, the system motion can neither grow unboundedly nor decay to zero. Instead, it displays a

sustained oscillation independent of initial conditions, as illustrated in Figure 1.4. This so-called

limit cycle is sustained by periodically releasing energy into and absorbing energy from the

environment, through the damping term. This is in contrast with the case of a conservative mass-

spring system, which does not exchange energy with its environment during its vibration. C3

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (23)

Sect. 1.2 Nonlinear System Behavior 9

Figure 1.4 : Responses of the Van der Pol oscillator

Of course, sustained oscillations can also be found in linear systems, in the caseof marginally stable linear systems (such as a mass-spring system without damping) orin the response to sinusoidal inputs. However, limit cycles in nonlinear systems aredifferent from linear oscillations in a number of fundamental aspects. First, theamplitude of the self-sustained excitation is independent of the initial condition, asseen in Figure 1.2, while the oscillation of a marginally stable linear system has itsamplitude determined by its initial conditions. Second, marginally stable linearsystems are very sensitive to changes in system parameters (with a slight changecapable of leading either to stable convergence or to instability), while limit cycles arenot easily affected by parameter changes.

Limit cycles represent an important phenomenon in nonlinear systems. Theycan be found in many areas of enginering and nature. Aircraft wing fluttering, a limitcycle caused by the interaction of aerodynamic forces and structural vibrations, isfrequently encountered and is sometimes dangerous. The hopping motion of a leggedrobot is another instance of a limit cycle. Limit cycles also occur in electrical circuits,e.g., in laboratory electronic oscillators. As one can see from these examples, limitcycles can be undesirable in some cases, but desirable in other cases. An engineer hasto know how to eliminate them when they are undesirable, and conversely how togenerate or amplify them when they are desirable. To do this, however, requires anunderstanding of the properties of limit cycles and a familiarity with the tools formanipulating them.

Bifurcations

As the parameters of nonlinear dynamic systems are changed, the stability of theequilibrium point can change (as it does in linear systems) and so can the number ofequilibrium points. Values of these parameters at which the qualitative nature of the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (24)

10 Introduction Chap. 1

system's motion changes are known as critical or bifurcation values. Thephenomenon of bifurcation, i.e., quantitative change of parameters leading toqualitative change of system properties, is the topic of bifurcation theory.

For instance, the smoke rising from an incense stick (smokestacks andcigarettes are old-fashioned) first accelerates upwards (because it is lighter than theambient air), but beyond some critical velocity breaks into swirls. More prosaically,let us consider the system described by the so-called undamped Duffing equation

x + ax + x3 = 0

(the damped Duffing equation is x + ex + ax + px3 = 0 , which may represent amass-damper-spring system with a hardening spring). We can plot the equilibriumpoints as a function of the parameter a. As a varies from positive to negative, oneequilibrium point splits into three points (xe = 0, "Noc, - ~ia ), as shown in Figure1.5(a). This represents a qualitative change in the dynamics and thus a = 0 is a criticalbifurcation value. This kind for bifurcation is known as a pitchfork, due to the shapeof the equilibrium point plot in Figure 1.5(a).

Another kind of bifurcation involves the emergence of limit cycles asparameters are changed. In this case, a pair of complex conjugate eigenvaluesPj = y + ja>, p2 = J-ju> cross from the left-half plane into the right-half plane, andthe response of the unstable system diverges to a limit cycle. Figure 1.5(b) depicts thechange of typical system state trajectories (states are x and x) as the parameter a isvaried. This type of bifurcation is called a Hopf bifurcation.

stable

stable

unstable

stable

(a)

Figure 1.5 : (a) a pitchfork bifurcation; (b) a Hopf bifurcation

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (25)

Sect. 1.2

Chaos

Nonlinear System Behavior 11

For stable linear systems, small differences in initial conditions can only cause smalldifferences in output. Nonlinear systems, however, can display a phenomenon calledchaos, by which we mean that the system output is extremely sensitive to initialconditions. The essential feature of chaos is the unpredictability of the system output.Even if we have an exact model of a nonlinear system and an extremely accuratecomputer, the system's response in the long-run still cannot be well predicted.

Chaos must be distinguished from random motion. In random motion, thesystem model or input contain uncertainty and, as a result, the time variation of theoutput cannot be predicted exactly (only statistical measures are available). In chaoticmotion, on the other hand, the involved problem is deterministic, and there is littleuncertainty in system model, input, or initial conditions.

As an example of chaotic behavior, let us consider the simple nonlinear system

= 6 sinr

which may represent a lightly-damped, sinusoidally forced mechanical structureundergoing large elastic deflections. Figure 1.6 shows the responses of the systemcorresponding to two almost identical initial conditions, namely x(0) = 2, x(0) = 3(thick line) and x(Q) = 2.01, x(0) = 3.01 (thin line). Due to the presence of the strongnonlinearity in x5, the two responses are radically different after some time.

-3.010 20 25 30 35 40 45 50

timefsec")

Figure 1.6 : Chaotic behavior of a nonlinear system

Chaotic phenomena can be observed in many physical systems. The mostcommonly seen physical problem is turbulence in fluid mechanics (such as the swirlsof our incense stick). Atmospheric dynamics also display clear chaotic behavior, thus

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (26)

12 Introduction Chap. 1

making long-term weather prediction impossible. Some mechanical and electricalsystems known to exhibit chaotic vibrations include buckled elastic structures,mechanical systems with play or backlash, systems with aeroelastic dynamics, wheel-rail dynamics in railway systems, and, of course, feedback control devices.

Chaos occurs mostly in strongly nonlinear systems. This implies that, for agiven system, if the initial condition or the external input cause the system to operatein a highly nonlinear region, it increases the possibility of generating chaos. Chaoscannot occur in linear systems. Corresponding to a sinusoidal input of arbitrarymagnitude, the linear system response is always a sinusoid of the same frequency. Bycontrast, the output of a given nonlinear system may display sinusoidal, periodic, orchaotic behaviors, depending on the initial condition and the input magnitude.

In the context of feedback control, it is of course of interest to know when anonlinear system will get into a chaotic mode (so as to avoid it) and, in case it does,how to recover from it. Such problems are the object of active research.

Other behaviors

Other interesting types of behavior, such as jump resonance, subharmonic generation,asynchronous quenching, and frequency-amplitude dependence of free vibrations, canalso occur and become important in some system studies. However, the abovedescription should provide ample evidence that nonlinear systems can haveconsiderably richer and more complex behavior than linear systems.

1.3 An Overview of the Book

Because nonlinear systems can have much richer and more complex behaviors thanlinear systems, their analysis is much more difficult. Mathematically, this is reflectedin two aspects. First, nonlinear equations, unlike linear ones, cannot in general besolved analytically, and therefore a complete understanding of the behavior of anonlinear system is very difficult. Second, powerful mathematical tools like Laplaceand Fourier transforms do not apply to nonlinear systems.

As a result, there are no systematic tools for predicting the behavior ofnonlinear systems, nor are there systematic procedures for designing nonlinear controlsystems. Instead, there is a rich inventory of powerful analysis and design tools, eachbest applicable to particular classes of nonlinear control problems. It is the objective ofthis book to present these various tools, with particular emphasis on their powers andlimitations, and on how they can be effectively combined.

This book is divided into two major parts. Part I (chapters 2-5) presents the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (27)

Sect. 1.4 Notes and References 13

major analytical tools that can be used to study a nonlinear system. Part II (chapters6-9) discusses the major nonlinear controller design techniques. Each part starts with ashort introduction providing the background for the main issues and techniques to bediscussed.

In chapter 2, we further familiarize ourselves with some basic nonlinear systembehaviors, by studying second-order systems using the simple graphical tools providedby so-called phase plane analysis. Chapter 3 introduces the most fundamental analysistool to be used in this book, namely the concept of a Lyapunov function and its use innonlinear stability analysis. Chapter 4 studies selected advanced topics in stabilityanalysis. Chapter 5 discusses an approximate nonlinear system analysis method, thedescribing function method, which aims at extending to nonlinear systems some of thedesirable and intuitive properties of linear frequency response analysis.

The basic idea of chapter 6 is to study under what conditions the dynamics of anonlinear system can be algebraically transformed in that of a linear system, on whichlinear control design techniques can in turn be applied. Chapters 7 and 8 then studyhow to reduce or practically eliminate the effects of model uncertainties on thestability and performance of feedback controllers for linear or nonlinear systems,using so-called robust and adaptive approaches. Finally, chapter 9 extensivelydiscusses the use of known physical properties to simplify and enhance the design ofcontrollers for complex multi-input nonlinear systems.

The book concentrates on nonlinear systems represented in continuous-timeform. Even though most control systems are implemented digitally, nonlinearphysical systems are continuous in nature and are hard to meaningfully discretize,while digital control systems may be treated as continuous-time systems in analysisand design if high sampling rates are used. Given the availability of cheapcomputation, the most common practical case when it may be advantageous toconsider sampling explicitly is when measurements are sparse, as e.g., in the case ofunderwater vehicles using acoustic navigation. Some practical issues involved in thedigital implementation of controllers designed from continuous-time formulations arediscussed in the introduction to Part II.

1.4 Notes and References

Detailed discussions of bifurcations and chaos can be found, e.g., in [Guckenheimer and Holmes,

1983] and in [Thompson and Stewart, 1986], from which the example of Figure 1.6 is adapted.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (28)

Part INonlinear Systems Analysis

The objective of this part is to present various tools available for analyzing nonlinearcontrol systems. The study of these nonlinear analysis techniques is important for anumber of reasons. First, theoretical analysis is usually the least expensive way ofexploring a system's characteristics. Second, simulation, though very important innonlinear control, has to be guided by theory. Blind simulation of nonlinear systems islikely to produce few results or misleading results. This is especially true given thegreat richness of behavior that nonlinear systems can exhibit, depending on initialconditions and inputs. Third, the design of nonlinear controllers is always based onanalysis techniques. Since design methods are usually based on analysis methods, it isalmost impossible to master the design methods without first studying the analysistools. Furthermore, analysis tools also allow us to assess control designs after theyhave been made, and, in case of inadequate performance, they may also suggestdirections of modifying the control designs.

It should not come as a surprise that no universal technique has been devised forthe analysis of all nonlinear control systems. In linear control, one can analyze asystem in the time domain or in the frequency domain. However, for nonlinearcontrol systems, none of these standard approaches can be used, since direct solutionof nonlinear differential equations is generally impossible, and frequency domaintransformations do not apply.

14

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (29)

Parti Nonlinear Systems Analysis 15

While the analysis of nonlinear control systems is difficult, serious efforts havebeen made to develop appropriate theoretical tools for it. Many methods of nonlinearcontrol system analysis have been proposed. Let us briefly describe some of thesemethods before discussing their details in the following chapters.

Phase plane analysis

Phase plane analysis, discussed in chapter 2, is a graphical method of studyingsecond-order nonlinear systems. Its basic idea is to solve a second order differentialequation graphically, instead of seeking an analytical solution. The result is a familyof system motion trajectories on a two-dimensional plane, called the phase plane,which allow us to visually observe the motion patterns of the system. While phaseplane analysis has a number of important advantages, it has the fundamentaldisadvantage of being applicable only to systems which can be well approximated bya second-order dynamics. Because of its graphical nature, it is frequently used toprovide intuitive insights about nonlinear effects.

Lyapunov theory

Basic Lyapunov theory comprises two methods introduced by Lyapunov, theindirect method and the direct method. The indirect method, or linearization method,states that the stability properties of a nonlinear system in the close vicinity of anequilibrium point are essentially the same as those of its linearized approximation. Themethod serves as the theoretical justification for using linear control for physicalsystems, which are always inherently nonlinear. The direct method is a powerful toolfor nonlinear system analysis, and therefore the so-called Lyapunov analysis oftenactually refers to the direct method. The direct method is a generalization of theenergy concepts associated with a mechanical system: the motion of a mechanicalsystem is stable if its total mechanical energy decreases all the time. In using thedirect method to analyze the stability of a nonlinear system, the idea is to construct ascalar energy-like function (a Lyapunov function) for the system, and to see whether itdecreases. The power of this method comes from its generality: it is applicable to allkinds of control systems, be they time-varying or time-invariant, finite dimensional orinfinite dimensional. Conversely, the limitation of the method lies in the fact that it isoften difficult to find a Lyapunov function for a given system.

Although Lyapunov's direct method is originally a method of stability analysis,it can be used for other problems in nonlinear control. One important application is thedesign of nonlinear controllers. The idea is to somehow formulate a scalar positivefunction of the system states, and then choose a control law to make this functiondecrease. A nonlinear control system thus designed will be guaranteed to be stable.Such a design approach has been used to solve many complex design problems, e.g.,

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (30)

16 Nonlinear Systems Analysis Part I

in robotics and adaptive control. The direct method can also be used to estimate theperformance of a control system and study its robustness. The important subject ofLyapunov analysis is studied in chapters 3 and 4, with chapter 3 presenting the mainconcepts and results in Lyapunov theory, and chapter 4 discussing some advancedtopics.

Describing functions

The describing function method is an approximate technique for studyingnonlinear systems. The basic idea of the method is to approximate the nonlinearcomponents in nonlinear control systems by linear "equivalents", and then usefrequency domain techniques to analyze the resulting systems. Unlike the phase planemethod, it is not restricted to second-order systems. Unlike Lyapunov methods,whose applicability to a specific system hinges on the success of a trial-and-errorsearch for a Lyapunov function, its application is straightforward for nonlinearsystems satisfying some easy-to-check conditions.

The method is mainly used to predict limit cycles in nonlinear systems. Otherapplications include the prediction of subharmonic generation and the determinationof system response to sinusoidal excitation. The method has a number of advantages.First, it can deal with low order and high order systems with the same straightforwardprocedure. Second, because of its similarity to frequency-domain analysis of linearsystems, it is conceptually simple and physically appealing, allowing users to exercisetheir physical and engineering insights about the control system. Third, it can dealwith the "hard nonlinearities" frequently found in control systems without anydifficulty. As a result, it is an important tool for practical problems of nonlinearcontrol analysis and design. The disadvantages of the method are linked to itsapproximate nature, and include the possibility of inaccurate predictions (falsepredictions may be made if certain conditions are not satisfied) and restrictions on thesystems to which it applies (for example, it has difficulties in dealing with systemswith multiple nonlinearities).

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (31)

Chapter 2Phase Plane Analysis

Phase plane analysis is a graphical method for studying second-order systems, whichwas introduced well before the turn of the century by mathematicians such as HenriPoincare. The basic idea of the method is to generate, in the state space of a second-order dynamic system (a two-dimensional plane called the phase plane), motiontrajectories corresponding to various initial conditions, and then to examine thequalitative features of the trajectories. In such a way, information concerning stabilityand other motion patterns of the system can be obtained. In this chapter, our objectiveis to gain familiarity with nonlinear systems through this simple graphical method.

Phase plane analysis has a number of useful properties. First, as a graphicalmethod, it allows us to visualize what goes on in a nonlinear system starting fromvarious initial conditions, without having to solve the nonlinear equations analytically.Second, it is not restricted to small or smooth nonlinearities, but applies equally wellto strong nonlinearities and to "hard" nonlinearities. Finally, some practical controlsystems can indeed be adequately approximated as second-order systems, and thephase plane method can be used easily for their analysis. Conversely, of course, thefundamental disadvantage of the method is that it is restricted to second-order (or first-order) systems, because the graphical study of higher-order systems iscomputationally and geometrically complex.

17

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (32)

18 Phase Plane Analysis Chap. 2

2.1 Concepts of Phase Plane Analysis

2.1.1 Phase Portraits

The phase plane method is concerned with the graphical study of second-orderautonomous systems described by

x2=f2(Xl,x2) (2.1b)

where jq and x2 are the states of the system, and/ , and/ 2 are nonlinear functions ofthe states. Geometrically, the state space of this system is a plane having x, and x2 ascoordinates. We will call this plane the phase plane.

Given a set of initial conditions x(0) = x0, Equation (2.1) defines a solutionx(0- With time / varied from zero to infinity, the solution x(t) can be representedgeometrically as a curve in the phase plane. Such a curve is called a phase planetrajectory. A family of phase plane trajectories corresponding to various initialconditions is called a phase portrait of a system.

To illustrate the concept of phase portrait, let us consider the following simplesystem.

Example 2.1: Phase portrait of a mass-spring system

The governing equation of the mass-spring system in Figure 2.1 (a) is the familiar linear second-

order differential equation

x+x = Q (2.2)

Assume that the mass is initially at rest, at length xo . Then the solution of the equation is

x(l) = xo cos t

x(t) = — A'osin(

Eliminating time / from the above equations, we obtain the equation of the trajectories

This represents a circle in the phase plane. Corresponding to different initial conditions, circles of

different radii can be obtained. Plotting these circles on the phase plane, we obtain a phase

portrait for the mass-spring system (Figure 2.1 .b). U

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (33)

Sect. 2.1 Concepts of Phase Plane Analysis 19

k= 1 m = l

(a) (b)

Figure 2.1 : A mass-spring system and its phase portrait

The power of the phase portrait lies in the fact that once the phase portrait of asystem is obtained, the nature of the system response corresponding to various initialconditions is directly displayed on the phase plane. In the above example, we easilysee that the system trajectories neither converge to the origin nor diverge to infinity.They simply circle around the origin, indicating the marginal nature of the system'sstability.

A major class of second-order systems can be described by differentialequations of the form

x +f(x, x) = 0

In state space form, this dynamics can be represented as

k\=x2

(2.3)

with A| = x and JT2 = -*• Most second-order systems in practice, such as mass-damper-spring systems in mechanics, or resistor-coil-capacitor systems in electricalengineering, can be represented in or transformed into this form. For these systems,the states are x and its derivative x. Traditionally, the phase plane method isdeveloped for the dynamics (2.3), and the phase plane is defined as the plane having xand x as coordinates. But it causes no difficulty to extend the method to more generaldynamics of the form (2.1), with the (xj , xj) plane as the phase plane, as we do in thischapter.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (34)

20 Phase Plane Analysis Chap. 2

2.1.2 Singular Points

An important concept in phase plane analysis is that of a singular point. A singularpoint is an equilibrium point in the phase plane. Since an equilibrium point is definedas a point where the system states can stay forever, this implies that x = 0, and using

(2.1),

/ , (* , , JC2) = 0 /2(jr1,jr2) = 0 (2.4)

The values of the equilibrium states can be solved from (2.4).

For a linear system, there is usually only one singular point (although in somecases there can be a continuous set of singular points, as in the system x + x = 0, forwhich all points on the real axis are singular points). However, a nonlinear systemoften has more than one isolated singular point, as the following example shows.

Example 2.2: A nonlinear second-order system

Consider the system

x + 0.6 x + 3 x + x1 = 0

whose phase portrait is plotted in Figure 2.2. The system has two singular points, one at (0, 0)

and the other at (-3, 0). The motion patterns of the system trajectories in the vicinity of the two

singular points have different natures. The trajectories move towards the point x = 0 while

moving away from the point x = — 3. D

One may wonder why an equilibrium point of a second-order system is called asingular point. To answer this, let us examine the slope of the phase trajectories.From (2.1), the slope of the phase trajectory passing through a point (X|,x2) isdetermined by

2 J2\ ! V ( 2 5 )

dx\ f\(xx,x2)

With the functions / ] and f2 assumed to be single valued, there is usually a definitevalue for this slope at any given point in phase plane. This implies that the phasetrajectories will not intersect. At singular points, however, the value of the slope is0/0, i.e., the slope is indeterminate. Many trajectories may intersect at such points, asseen from Figure 2.2. This indeterminacy of the slope accounts for the adjective"singular".

Singular points are very important features in the phase plane. Examination ofthe singular points can reveal a great deal of information about the properties of a

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (35)

Sect. 2.1 Concepts of Phase Plane Analysis 21

to infinity

Figure 2.2 : The phase portrait of a nonlinear system

system. In fact, the stability of linear systems is uniquely characterized by the natureof their singular points. For nonlinear systems, besides singular points, there may bemore complex features, such as limit cycles. These issues will be discussed in detailin sections 2.3 and 2.4.

Note that, although the phase plane method is developed primarily for second-order systems, it can also be applied to the analysis of first-order systems of the form

x +f(x) = 0

The idea is still to plot x with respect to x in the phase plane. The difference now isthat the phase portrait is composed of a single trajectory.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (36)

22 Phase Plane Analysis

Example 2.3: A first-order system

Consider the system

Chap. 2

There are three singular points, defined by - 4x + x 3 = 0, namely, x = 0, - 2 , and 2. The phase-

portrait of the system consists of a single trajectory, and is shown in Figure 2.3. The arrows in

the figure denote the direction of motion, and whether they point toward the left or the right at a

particular point is determined by the sign of x at that point. It is seen from the phase portrait of

this system that the equilibrium point x = 0 is stable, while the other two are unstable. O

stable

unstableFigure 2.3 : Phase trajectory of a first-

order system

2.1.3 Symmetry in Phase Plane Portraits

A phase portrait may have a priori known symmetry properties, which can simplify itsgeneration and study. If a phase portrait is symmetric with respect to the X\ or the x2

axis, one only needs in practice to study half of it. If a phase portrait is symmetricwith respect to both the Xj and x2 axes, only one quarter of it has to be explicitlyconsidered.

Before generating a phase portrait itself, we can determine its symmetryproperties by examining the system equations. Let us consider the second-orderdynamics (2.3). The slope of trajectories in the phase plane is of the form

dx2 f{x\,x2)dx,1

Since symmetry of the phase portraits also implies symmetry of the slopes (equal inabsolute value but opposite in sign), we can identify the following situations:

Symmetry about the xi axis: The condition is

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (37)

Sect. 2.2 Constructing Phase Portraits 23

f(xhx2) = f(xl,-x2)

This implies that the function / should be even in x2- The mass-spring system inExample 2.1 satisfies this condition. Its phase portrait is seen to be symmetric about

axis.

Symmetry about the x2 axis: Similarly,

f(x\,x2) = -f(-xl,x2)

implies symmetry with respect to the x2 axis. The mass-spring system also satisfiesthis condition.

Symmetry about the origin: When

f{x{,x2) = -f(-xh-x2)

the phase portrait of the system is symmetric about the origin.

2.2 Constructing Phase Portraits

Today, phase portraits are routinely computer-generated. In fact, it is largely theadvent of the computer in the early 1960's, and the associated ease of quicklygenerating phase portraits, which spurred many advances in the study of complexnonlinear dynamic behaviors such as chaos. However, of course (as e.g., in the case ofroot locus for linear systems), it is still practically useful to learn how to roughlysketch phase portraits or quickly verify the plausibility of computer outputs.

There are a number of methods for constructing phase plane trajectories forlinear or nonlinear systems, such as the so-called analytical method, the method ofisoclines, the delta method, Lienard's method, and Pell's method. We shall discusstwo of them in this section, namely, the analytical method and the method of isoclines.These methods are chosen primarily because of their relative simplicity. Theanalytical method involves the analytical solution of the differential equationsdescribing the systems. It is useful for some special nonlinear systems, particularlypiece-wise linear systems, whose phase portraits can be constructed by piecingtogether the phase portraits of the related linear systems. The method of isoclines is agraphical method which can conveniently be applied to construct phase portraits forsystems which cannot be solved analytically, which represent by far the most commoncase.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (38)

24 Phase Plane Analysis Chap. 2

ANALYTICAL METHOD

There are two techniques for generating phase plane portraits analytically. Bothtechniques lead to a functional relation between the two phase variables Xj and x2 inthe form

g(xhx2,c) = 0 (2.6)

where the constant c represents the effects of initial conditions (and, possibly, ofexternal input signals). Plotting this relation in the phase plane for different initialconditions yields a phase portrait.

The first technique involves solving equations (2.1) forx[ and x2 as functions oftime t, i.e.,

and then eliminating time t from these equations, leading to a functional relation in theform of (2.6). This technique was already illustrated in Example 2.1.

The second technique, on the other hand, involves directly eliminating the timevariable, by noting that

and then solving this equation for a functional relation between Xj and x2. Let us usethis technique to solve the mass-spring equation again.

Example 2.4: Mass-spring system

By noting that x = (dx/dx)(dx/dt), we can rewrite (2.2) as

-v — + x = 0dx

Integration of this equation yields

i 2 + x 2 =xo2 •

One sees that the second technique is more straightforward in generating the equationsfor the phase plane trajectories.

Most nonlinear systems cannot be easily solved by either of the above twotechniques. However, for piece-wise linear systems, an important class of nonlinearsystems, this method can be conveniently used, as the following example shows.

L

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (39)

Sect. 2.2

Example 2.5: A satellite control system

Constructing Phase Portraits 25

Figure 2.4 shows the control system for a simple satellite model. The satellite, depicted in Figure

2.5(a), is simply a rotational unit inertia controlled by a pair of thrusters, which can provide either

a positive constant torque U (positive firing) or a negative torque — U (negative firing). The

purpose of the control system is to maintain the satellite antenna at a zero angle by appropriately

firing the thrusters. The mathematical model of the satellite is

where w is the torque provided by the thrusters and 8 is the satellite angle.

Jets Satellite

ed = o U' —

-u

i u

1p

eip

Figure 2.4 : Satellite control system

Let us examine on the phase plane the behavior of the control system when the thrusters are

fired according to the control law

u(t) = / - U if 9 > 0w 1 u if e < o

(2.7)

which means that the thrusters push in the counterclockwise direction if G is positive, and vice

versa.

As the first step of the phase portrait generation, let us consider the phase portrait when the

thrusters provide a positive torque U. The dynamics of the system is

which implies that 6 dQ = U dQ. Therefore, the phase trajectories are a family of parabolas

defined by

where cf is a constant. The corresponding phase portrait of the system is shown in Figure 2.5(b).

When the thrusters provide a negative torque - U, the phase trajectories are similarly found

to be

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (40)

26 Phase Plane Analysis Chap. 2

u = -U

(a) (b) (c)

Figure 2.5 : Satellite control using on-off thrusters

with the corresponding phase portrait shown in Figure 2.5(c).

parabolictrajectories

u = +U

switching line

Figure 2.6 : Complete phase portrait of the control system

The complete phase portrait of the closed-loop control system can be obtained simply by

connecting the trajectories on the left half of the phase plane in 2.5(b) with those on the right half

of the phase plane in 2.5(c), as shown in Figure 2.6. The vertical axis represents a switching line,

because the control input and thus the phase trajectories are switched on that line. It is interesting

to see that, starting from a nonzero initial angle, the satellite will oscillate in periodic motions

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (41)

Sect. 2.2 Constructing Phase Portraits 27

under the action of the jets. One concludes from this phase portrait that the system is marginally

stable, similarly to the mass-spring system in Example 2.1. Convergence of the system to the

zero angle can be obtained by adding rate feedback (Exercise 2.4). [3

THE METHOD OF ISOCLINES

The basic idea in this method is that of isoclines. Consider the dynamics in (2.1). At apoint (JCJ , x2) in the phase plane, the slope of the tangent to the trajectory can bedetermined by (2.5). An isocline is defined to be the locus of the points with a giventangent slope. An isocline with slope a is thus defined to be

dx2 _f2(xh x2) _

dxx fl(xl,x2)

This is to say that points on the curve

all have the same tangent slope a.

In the method of isoclines, the phase portrait of a system is generated in twosteps. In the first step, a field of directions of tangents to the trajectories is obtained. Inthe second step, phase plane trajectories are formed from the field of directions .

Let us explain the isocline method on the mass-spring system in (2.2). Theslope of the trajectories is easily seen to be

dx2 X\

dx\ x2

Therefore, the isocline equation for a slope a is

X| + ax2 =0

i.e., a straight line. Along the line, we can draw a lot of short line segments with slopea. By taking a to be different values, a set of isoclines can be drawn, and a field ofdirections of tangents to trajectories are generated, as shown in Figure 2.7. To obtaintrajectories from the field of directions, we assume that the the tangent slopes arelocally constant. Therefore, a trajectory starting from any point in the plane can befound by connecting a sequence of line segments.

Let us use the method of isoclines to study the Van der Pol equation, anonlinear equation.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (42)

28 Phase Plane Analysis Chap. 2

Figure 2.7 : Isoclines for the mass-spring

system

Example 2.6: The Van der Pol equation

For the Van der Pol equation

an isocline of slope a is defined by

dx_0.2(x2- \)x + x

Therefore, the points on the curve

0 . 2 ( x 2 - \)x + x + ax = 0

all have the same slope a.

By taking a of different values, different isoclines can be obtained, as plotted in Figure 2.8.

Short line segments are drawn on the isoclines to generate a field of tangent directions. The phase

portraits can then be obtained, as shown in the plot. It is interesting to note that there exists a

closed curve in the portrait, and the trajectories starting from both outside and inside converge to

this curve. This closed curve corresponds to a limit cycle, as will be discussed further in section

2.5. •

Note that the same scales should be used for the xj axis and Xj axis of the phase

plane, so that the derivative dx^dx-^ equals the geometric slope of the trajectories.

Also note that, since in the second step of phase portrait construction we essentially

assume that the slope of the phase plane trajectories is locally constant, more isoclines

should be plotted in regions where the slope varies quickly, to improve accuracy.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (43)

Sect. 2.3 Determining Time from Phase Portraits 29

a = -5

a = l

trajectory

isoclines

Figure 2.8 : Phase portrait of the Van der Pol equation

2.3 Determining Time from Phase Portraits

Note that time t does not explicitly appear in the phase plane having Xy and x2 ascoordinates. However, in some cases, we might be interested in the time information.For example, one might want to know the time history of the system states startingfrom a specific initial point. Another relevant situation is when one wants to knowhow long it takes for the system to move from a point to another point in a phase planetrajectory. We now describe two techniques for computing time history from phaseportraits. Both techniques involve a step-by step procedure for recovering time.

Obtaining time from At~Ax/x

In a short time At, the change of x is approximately

Ax ~ xAt (2.8)

where x is the velocity corresponding to the increment Ax. Note that for a Ax of finitemagnitude, the average value of velocity during a time increment should be used toimprove accuracy. From (2.8), the length of time corresponding to the increment Ax

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (44)

30 Phase Plane Analysis Chap. 2

is

The above reasoning implies that, in order to obtain the time corresponding to themotion from one point to another point along a trajectory, one should divide thecorresponding part of the trajectory into a number of small segments (not necessarilyequally spaced), find the time associated with each segment, and then add up theresults. To obtain the time history of states corresponding to a certain initialcondition, one simply computes the time t for each point on the phase trajectory, andthen plots x with respect to t and x with respect to t,

Obtaining time from t = f (1/i) dx

Since x = dx/dt, we can write dt - dx/x. Therefore,

where x corresponds to time t and xo corresponds to time t0 . This equation impliesthat, if we plot a phase plane portrait with new coordinates x and (1/i), then the areaunder the resulting curve is the corresponding time interval.

2.4 Phase Plane Analysis of Linear Systems

In this section, we describe the phase plane analysis of linear systems. Besidesallowing us to visually observe the motion patterns of linear systems, this will alsohelp the development of nonlinear system analysis in the next section, because anonlinear systems behaves similarly to a linear system around each equilibrium point.

The general form of a linear second-order system is

xl=axr+ bx2 (2.9a)

k2 = cxi+dx2 (2.9b)

To facilitate later discussions, let us transform this equation into a scalar second-orderdifferential equation. Note from (2.9a) and (2.9b) that

b k2 = b cx\ + d(x\ — axj)

Consequently, differentiation of (2.9a) and then substitution of (2.9b) leads to

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (45)

Sect. 2.4 Phase Plane Analysis of Linear Systems 31

Xj = (a +d)X\ + (cb - ad)xi

Therefore, we will simply consider the second-order linear system described by

x + ax + bx = 0 (2.10)

To obtain the phase portrait of this linear system, we first solve for the timehistory

x(t) = klexit + k2e

l2> forX,*^2 (2.11a)

x(t) = klexi' + k2tehl for X{ = X^ (2.11b)

where the constants X\ and X2 are the solutions of the characteristic equation

s2 + as + b = (s - A,j) (s - Xj) =0

The roots A,j and > can be explicitly represented as

For linear systems described by (2.10), there is only one singular point (assumingb & 0), namely the origin. However, the trajectories in the vicinity of this singularitypoint can display quite different characteristics, depending on the values of a and b.The following cases can occur

1. ^.j and Xj are both real and have the same sign (positive or negative)

2. X\ and Xj are both real and have opposite signs

3. A,j and X2 are complex conjugate with non-zero real parts

4. X{ and X2 are complex conjugates with real parts equal to zero

We now briefly discuss each of the above four cases.

STABLE OR UNSTABLE NODE

The first case corresponds to a node. A node can be stable or unstable. If theeigenvalues are negative, the singularity point is called a stable node because both x(f)and x(t) converge to zero exponentially, as shown in Figure 2.9(a). If botheigenvalues are positive, the point is called an unstable node, because both x(t) andx{t) diverge from zero exponentially, as shown in Figure 2.9(b). Since the eigenvaluesare real, there is no oscillation in the trajectories.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (46)

32 Phase Plane Analysis Chap. 2

SADDLE POINT

The second case (say X^ < 0 and A > 0) corresponds to a saddle point (Figure 2.9(c)).The phase portrait of the system has the interesting "saddle" shape shown in Figure2.9(c). Because of the unstable pole Xj , almost all of the system trajectories divergeto infinity. In this figure, one also observes two straight lines passing through theorigin. The diverging line (with arrows pointing to infinity) corresponds to initialconditions which make £2 (i.e., the unstable component) equal zero. The convergingstraight line corresponds to initial conditions which make kl equal zero.

STABLE OR UNSTABLE FOCUS

The third case corresponds to a focus. A stable focus occurs when the real part of theeigenvalues is negative, which implies that x(t) and x(t) both converge to zero. Thesystem trajectories in the vicinity of a stable focus are depicted in Figure 2.9(d). Notethat the trajectories encircle the origin one or more times before converging to it,unlike the situation for a stable node. If the real part of the eigenvalues is positive,then x(t) and x(t) both diverge to infinity, and the singularity point is called anunstable focus. The trajectories corresponding to an unstable focus are sketched inFigure 2.9(e).

CENTER POINT

The last case corresponds to a center point, as shown in Figure 2.9(f). The namecomes from the fact that all trajectories are ellipses and the singularity point is thecenter of these ellipses. The phase portrait of the undamped mass-spring systembelongs to this category.

Note that the stability characteristics of linear systems are uniquely determinedby the nature of their singularity points. This, however, is not true for nonlinearsystems.

2.5 Phase Plane Analysis of Nonlinear Systems

In discussing the phase plane analysis of nonlinear systems, two points should be keptin mind. Phase plane analysis of nonlinear systems is related to that of linear systems,because the local behavior of a nonlinear system can be approximated by the behaviorof a linear system. Yet, nonlinear systems can display much more complicatedpatterns in the phase plane, such as multiple equilibrium points and limit cycles. Wenow discuss these points in more detail.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (47)

Sect. 2.5 Phase Plane Analysis of Nonlinear Systems 33

stable node

(a)

11

unstable nodeX X - C7

(b)77

saddle point

(c)

stable focus

(d)

unstable focus

x

(e)

center point

(0

Figure 2.9 : Phase-portraits of linear systems

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (48)

34 Phase Plane Analysis Chap. 2

LOCAL BEHAVIOR OF NONLINEAR SYSTEMS

In the phase portrait of Figure 2.2, one notes that, in contrast to linear systems, thereare two singular points, (0,0) and (-3,0) . However, we also note that the features ofthe phase trajectories in the neighborhood of the two singular points look very muchlike those of linear systems, with the first point corresponding to a stable focus and thesecond to a saddle point. This similarity to a linear system in the local region of eachsingular point can be formalized by linearizing the nonlinear system, as we nowdiscuss.

If the singular point of interest is not at the origin, by defining the differencebetween the original state and the singular point as a new set of state variables, onecan always shift the singular point to the origin. Therefore, without loss of generality,we may simply consider Equation (2.1) with a singular point at 0. Using Taylorexpansion, Equations (2.1a) and (2.1b) can be rewritten as

h = c x l + dx2 + 82^1 ' X2>

where gj and g2 contain higher order terms.

In the vicinity of the origin, the higher order terms can be neglected, andtherefore, the nonlinear system trajectories essentially satisfy the linearized equation

JL'j = axl + bx2

x2 = cxi+dx2

As a result, the local behavior of the nonlinear system can be approximated by thepatterns shown in Figure 2.9.

LIMIT CYCLES

m the phase portrait of the nonlinear Van der Pol equation, shown in Figure 2.8, oneobserves that the system has an unstable node at the origin. Furthermore, there is aclosed curve in the phase portrait. Trajectories inside the curve and those outside thecurve all tend to this curve, while a motion started on this curve will stay on it forever,circling periodically around the origin. This curve is an instance of the so-called"limit cycle" phenomenon. Limit cycles are unique features of nonlinear systems.

In the phase plane, a limit cycle is defined as an isolated closed curve. Thetrajectory has to be both closed, indicating the periodic nature of the motion, andisolated, indicating the limiting nature of the cycle (with nearby trajectories

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (49)

Sect. 2.5 Phase Plane Analysis of Nonlinear Systems 35

converging or diverging from it). Thus, while there are many closed curves in thephase portraits of the mass-spring-damper system in Example 2.1 or the satellitesystem in Example 2.5, these are not considered limit cycles in this definition, becausethey are not isolated.

Depending on the motion patterns of the trajectories in the vicinity of the limitcycle, one can distinguish three kinds of limit cycles

1. Stable Limit Cycles: all trajectories in the vicinity of the limit cycleconverge to it as t —> °° (Figure 2.10(a));

2. Unstable Limit Cycles: all trajectories in the vicinity of the limit cyclediverge from it as t -> °° (Figure 2.10(b));

3. Semi-Stable Limit Cycles: some of the trajectories in the vicinityconverge to it, while the others diverge from it as r —» °° (Figure

2 divergingconverging diverging

(a) (b) (c)

Figure 2.10 : Stable, unstable, and semi-stable limit cycles

As seen from the phase portrait of Figure 2.8, the limit cycle of the Van der Polequation is clearly stable. Let us consider some additional examples of stable,unstable, and semi-stable limit cycles.

Example 2.7: stable, unstable, and semi-stable limit cycles

Consider the following nonlinear systems

(a)

(b)

(c)

l=x2-xx(xl

X , = J

- I)2

-x2(xf

+x2(x,2

-x2(x,2

+ x 2 - -

+ x 22 -

+ x 22 -

1)

I ) 2

(2.12)

(2.13)

(2.14)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (50)

36 Phase Plane Analysis Chap. 2

Let us study system (a) first. By introducing polar coordinates

/• = (x 12 + x2

2)1/2 9 = tan-1(jc2/x1)

the dynamic equations (2.12) are transformed as

dr , , , . d<d

T<=-r(r-l) Tr-X

When the state starts on the unit circle, the above equation shows that r(t) = 0. Therefore, the state

will circle around the origin with a period 1/2K. When r < 1, then r > 0. This implies that the state

tends to the circle from inside. When r > 1, then /• < 0. This implies that the state tends toward

the unit circle from outside. Therefore, the unit circle is a stable limit cycle. This can also be

concluded by examining the analytical solution of (2.12)

r(t) = 1 6(0 = Qn - 1( l+c oe- 2 ' ) 1 / 2

where

Similarly, one can find that the system (b) has an unstable limit cycle and system (c) has a semi-

stable limit cycle. Q

2.6 Existence of Limit Cycles

As mentioned in chapter 1, it is of great importance for control engineers to predict theexistence of limit cycles in control systems. In this section, we state three simpleclassical theorems to that effect. These theorems are easy to understand and apply.

The first theorem to be presented reveals a simple relationship between theexistence of a limit cycle and the number of singular points it encloses. In thestatement of the theorem, we use N to represent the number of nodes, centers, and focienclosed by a limit cycle, and S to represent the number of enclosed saddle points.

Theorem 2.1 (Poincare) / / a limit cycle exists in the second-order autonomoussystem (2.1), then N = S + 1 .

This theorem is sometimes called the index theorem. Its proof is mathematicallyinvolved (actually, a family of such proofs led to the development of algebraictopology) and shall be omitted here. One simple inference from this theorem is that alimit cycle must enclose at least one equilibrium point. The theorem's result can be

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (51)

Sect. 2.6 Existence of Limit Cycles 37

verified easily on Figures 2.8 and 2.10.

The second theorem is concerned with the asymptotic properties of thetrajectories of second-order systems.

Theorem 2.2 (Poincare-Bendixson) If a trajectory of the second-orderautonomous system remains in a finite region Q, then one of the following is true:

(a) the trajectory goes to an equilibrium point

(b) the trajectory tends to an asymptotically stable limit cycle

(c) the trajectory is itself a limit cycle

While the proof of this theorem is also omitted here, its intuitive basis is easy to see,and can be verified on the previous phase portraits.

The third theorem provides a sufficient condition for the non-existence of limitcycles.

Theorem 2.3 (Bendixson) For the nonlinear system (2.1), no limit cycle can existin a region Q. of the phase plane in which 3/j /3xj + 3/2/3.X2 does not vanish anddoes not change sign.

Proof: Let us prove this theorem by contradiction. First note that, from (2.5), the equation

0 (2.15)

is satisfied for any system trajectories, including a limit cycle. Thus, along the closed curve L of

a limit cycle, we have

f (/,rfjc2-/2rfx-1> = 0 (2.16)

Using Stokes' Theorem in calculus, we have

where the integration on the right-hand side is carried out on the area enclosed by the limit cycle.

By Equation (2.16), the left-hand side must equal zero. This, however, contradicts the fact

that the right-hand side cannot equal zero because by hypothesis 3/j/3xj +3/2 /3x2 does not

vanish and does not change sign. El

Let us illustrate the result on an example.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (52)

38 Phase Plane Analysis Chap. 2

Example 2.8: Consider the nonlinear system

x2 =

Since

which is always strictly positive (except at the origin), the system does not have any limit cycles

anywhere in the phase plane. . \3

The above three theorems represent very powerful results. It is important tonotice, however, that they have no equivalent in higher-order systems, where exoticasymptotic behaviors other than equilibrium points and limit cycles can occur.

2.7 Summary

Phase plane analysis is a graphical method used to study second-order dynamicsystems. The major advantage of the method is that it allows visual examination of theglobal behavior of systems. The major disadvantage is that it is mainly limited tosecond-order systems (although extensions to third-order systems are often achievedwith the aid of computer graphics). The phenomena of multiple equilibrium points andof limit cycles are clearly seen in phase plane analysis. A number of useful classicaltheorems for the prediction of limit cycles in second-order systems are also presented.

2.8 Notes and References

Phase plane analysis is a very classical topic which has been addressed by numerous control texts.

An extensive treatment can be found in [Graham and McRuer, 1961]. Examples 2.2 and 2.3 are

adapted from [Ogata, 1970]. Examples 2.5 and 2.6 and section 2.6 are based on [Hsu and Meyer,

1968].

2.9 Exercises

2.1 Draw the phase portrait and discuss the properties of the linear, unity feedback control system

of open-loop transfer function

1 0

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (53)

Sect. 2.9

2.2 Draw the phase portraits of the following systems, using isoclines

(a) e + e + 0.5 e = o

(b) e + e + o.5 e = i

Exercises 39

2.3 Consider the nonlinear system

x = y + x(x* + yl- 1) sin

y = - x + y (x2 + y2 - 1) sin

Without solving the above equations explicitly, show that the system has infinite number of limit

cycles. Determine the stability of these limit cycles. (Hint: Use polar coordinates.)

2.4 The system shown in Figure 2.10 represents a satellite control system with rate feedback

provided by a gyroscope. Draw the phase portrait of the system, and determine the system's

stability.

p + a -1 'u 1

P1

Figure 2.10 : Satellite control system with rate feedback

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (54)

Chapter 3Fundamentals of Lyapunov Theory

Given a control system, the first and most important question about its variousproperties is whether it is stable, because an unstable control system is typicallyuseless and potentially dangerous. Qualitatively, a system is described as stable ifstarting the system somewhere near its desired operating point implies that it will stayaround the point ever after. The motions of a pendulum starting near its twoequilibrium points, namely, the vertical up and down positions, are frequently used toillustrate unstable and stable behavior of a dynamic system. For aircraft controlsystems, a typical stability problem is intuitively related to the following question:will a trajectory perturbation due to a gust cause a significant deviation in the laterflight trajectory? Here, the desired operating point of the system is the flighttrajectory in the absence of disturbance. Every control system, whether linear ornonlinear, involves a stability problem which should be carefully studied.

The most useful and general approach for studying the stability of nonlinearcontrol systems is the theory introduced in the late 19th century by the Russianmathematician Alexandr Mikhailovich Lyapunov. Lyapunov's work, The GeneralProblem of Motion Stability, includes two methods for stability analysis (the so-calledlinearization method and direct method) and was first published in 1892. Thelinearization method draws conclusions about a nonlinear system's local stabilityaround an equilibrium point from the stability properties of its linear approximation.The direct method is not restricted to local motion, and determines the stabilityproperties of a nonlinear system by constructing a scalar "energy-like" function for thesystem and examining the function's time variation. For over half a century, however,

40

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (55)

Sect. 3.1 Nonlinear Systems and Equilibrium Points 41

Lyapunov's pioneering work on stability received little attention outside Russia,although it was translated into French in 1908 (at the instigation of Poincare), andreprinted by Princeton University Press in 1947. The publication of the work of Lur'eand a book by La Salle and Lefschetz brought Lyapunov's work to the attention of thelarger control engineering community in the early 1960's. Many refinements ofLyapunov's methods have since been developed. Today, Lyapunov's linearizationmethod has come to represent the theoretical justification of linear control, whileLyapunov's direct method has become the most important tool for nonlinear systemanalysis and design. Together, the linearization method and the direct methodconstitute the so-called Lyapunov stability theory.

The objective of this and the next chapter is to present Lyapunov stabilitytheory and illustrate its use in the analysis and the design of nonlinear systems. Toprevent mathematical complexity from obscuring the theoretical concepts, this chapterpresents the most basic results of Lyapunov theory in terms of autonomous {i.e., time-invariant) systems, leaving more advanced topics to chapter 4. This chapter isorganized as follows. In section 3.1, we provide some background definitionsconcerning nonlinear systems and equilibrium points. In section 3.2, various conceptsof stability are described to characterize different aspects of system behavior.Lyapunov's linearization method is presented in section 3.3. The most useful theoremsin the direct method are studied in section 3.4. Section 3.5 is devoted to the questionof how to use these theorems to study the stability of particular classes of nonlinearsystems. Section 3.6 sketches how the direct method can be used as a powerful wayof designing controllers for nonlinear systems.

3.1 Nonlinear Systems and Equilibrium Points

Before addressing the main problems of defining and determining stability in the nextsections, let us discuss some relatively simple background issues.

NONLINEAR SYSTEMS

A nonlinear dynamic system can usually be represented by a set of nonlineardifferential equations in the form

x = f(x,r) (3.1)

where f is a «xl nonlinear vector function, and x is the nxl state vector. A particularvalue of the state vector is also called a point because it corresponds to a point in thestate-space. The number of states n is called the order of the system. A solution x(0of the equations (3.1) usually corresponds to a curve in state space as t varies from

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (56)

42 Fundamentals ofLyapunov Theory Chap. 3

zero to infinity, as already seen in phase plane analysis for the case n = 2. This curve isgenerally referred to as a state trajectory or a system trajectory.

It is important to note that although equation (3.1) does not explicitly containthe control input as a variable, it is directly applicable to feedback control systems.The reason is that equation (3.1) can represent the closed-loop dynamics of a feedbackcontrol system, with the control input being a function of state x and time /, andtherefore disappearing in the closed-loop dynamics. Specifically, if the plant dynamics

x = f(x, u,0

and some control law has been selected

u = g(x, t)

then the closed-loop dynamics is

x = f[x,g(x,r),r]

which can be rewritten in the form (3.1). Of course, equation (3.1) can also representdynamic systems where no control signals are involved, such as a freely swingingpendulum.

A special class of nonlinear systems are linear systems. The dynamics of linearsystems are of the form

i = A(f)x

where A(t) is an nxn matrix.

AUTONOMOUS AND NON-AUTONOMOUS SYSTEMS

Linear systems are classified as either time-varying or time-invariant, depending onwhether the system matrix A varies with time or not. In the more general context ofnonlinear systems, these adjectives are traditionally replaced by "autonomous" and"non-autonomous".

Definition 3.1 The nonlinear system (3.1) is said to be autonomous if f does notdepend explicitly on time, i.e., if the system's state equation can be written

x = f (x) (3.2)

Otherwise, the system is called non-autonomous .

Obviously, linear time-invariant (LTI) systems are autonomous and linear time-

I

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (57)

Sect. 3.1 Nonlinear Systems and Equilibrium Points 43

varying (LTV) systems are non-autonomous. The second-order systems studied inchapter 2 are all autonomous.

Strictly speaking, all physical systems are non-autonomous, because none oftheir dynamic characteristics is strictly time-invariant. The concept of an autonomoussystem is an idealized notion, like the concept of a linear system. In practice,however, system properties often change very slowly, and we can neglect their timevariation without causing any practically meaningful error.

It is important to note that for control systems, the above definition is made onthe closed-loop dynamics. Since a control system is composed of a controller and aplant (including sensor and actuator dynamics), the non-autonomous nature of acontrol system may be due to a time-variation either in the plant or in the control law.Specifically, a time-invariant plant with dynamics

x = f (x, u)

may lead to a non-autonomous closed-loop system if a controller dependent on time tis chosen, i.e., if u = g(x, f). For example, the closed-loop system of the simple plantx = - x + u can be nonlinear and non-autonomous by choosing u to be nonlinear andtime-varying (e.g., u = -x2 sin t). In fact, adaptive controllers for linear time-invariantplants usually make the closed-loop control systems nonlinear and non-autonomous.

The fundamental difference between autonomous and non-autonomous systemslies in the fact that the state trajectory of an autonomous system is independent of theinitial time, while that of a non-autonomous system generally is not. As we will see inthe next chapter, this difference requires us to consider the initial time explicitly indefining stability concepts for non-autonomous systems, and makes the analysis moredifficult than that of autonomous systems.

It is well known that the analysis of linear time-invariant systems is much easierthan that of linear time-varying systems. The same is true with nonlinear systems.Generally speaking, autonomous systems have relatively simpler properties and theiranalysis is much easier. For this reason, in the remainder of this chapter, we willconcentrate on the analysis of autonomous systems, represented by (3.2). Extensionsof the concepts and results to non-autonomous systems will be studied in chapter 4.

EQUILIBRIUM POINTS

It is possible for a system trajectory to correspond to only a single point. Such a pointis called an equilibrium point. As we shall see later, many stability problems arenaturally formulated with respect to equilibrium points.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (58)

44 Fundamentals ofLyapunov Theory Chap. 3

Definition 3.2 A state x* is an equilibrium state (or equilibrium point) of the systemif once x(t) is equal to x , it remains equal to x for all future time.

Mathematically, this means that the constant vector x* satisfies

0 = f(x*) (3.3)

Equilibrium points can be found by solving the nonlinear algebraic equations (3.3).

A linear time-invariant system

x = Ax (3.4)

has a single equilibrium point (the origin 0) if A is nonsingular. If A is singular, it hasan infinity of equilibrium points, which are contained in the null-space of the matrixA, i.e., the subspace defined by Ax = 0. This implies that the equilibrium points arenot isolated, as reflected by the example x + x = 0 , for which all points on the x axisof the phase plane are equilibrium points.

A nonlinear system can have several (or infinitely many) isolated equilibriumpoints, as seen in Example 1.1. The following example involves a familiar physicalsystem.

Example 3.1: The Pendulum

Consider the pendulum of Figure 3.1, whose dynamics is given by the following nonlinear

autonomous equation

MR2 8 + b 6 + MgR sin 8 = 0 (3.5)

Figure 3.1 : The pendulum

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (59)

Sect. 3.1 Nonlinear Systems and Equilibrium Points 45

where R is the pendulum's length, M its mass, b the friction coefficient at the hinge, and g the

gravity constant. Letting x^ = 8 , x2 = 0, the corresponding state-space equation is

x , = x2 (3.6a)

b sx-y = x9 - — sinx, (3.6b)

MR1 R

Therefore, the equilibrium points are given by

x2 = 0 , sin X| = 0

which leads to the points (0 [2ji], 0) and (JI [27i], 0). Physically, these points correspond to the

pendulum resting exactly at the vertical up and down positions. Q

In linear system analysis and design, for notational and analytical simplicity, weoften transform the linear system equations in such a way that the equilibrium point isthe origin of the state-space. We can do the same thing for nonlinear systems (3.2),about a specific equilibrium point. Let us say that the equilibrium point of interest isx*. Then, by introducing a new variable

y = x-x*

and substituting x = y + x into equations (3.2), a new set of equations on the variabley are obtained

y = f(y+x*) (3.7)

One can easily verify that there is a one-to-one correspondence between the solutionsof (3.2) and those of (3.7), and that in addition, y=0, the solution corresponding tox = x*, is an equilibrium point of (3.7). Therefore, instead of studying the behavior ofthe equation (3.2) in the neighborhood of x*, one can equivalently study the behaviorof the equations (3.7) in the neighborhood of the origin.

NOMINAL MOTION

In some practical problems, we are not concerned with stability around an equilibriumpoint, but rather with the stability of a motion, i.e, whether a system will remain closeto its original motion trajectory if slightly perturbed away from it, as exemplified bythe aircraft trajectory control problem mentioned at the beginning of this chapter. Wecan show that this kind of motion stability problem can be transformed into anequivalent stability problem around an equilibrium point, although the equivalentsystem is now non-autonomous.

Let \*(t) be the solution of equation (3.2), i.e., the nominal motion trajectory,corresponding to initial condition x*(0) = x0. Let us now perturb the initial condition

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (60)

46 Fundamentals of Lyapunov Theory Chap. 3

to be x(0) = xo + 8x0 and study the associated variation of the motion error

e(f) = x(r) - x*(0

as illustrated in Figure 3.2. Since both x*(t) and x(/) are solutions of (3.2), we have

I

Figure 3.2 : Nominal and Perturbed Motions

x* = f (x*) x(0) = x0

x = f (x) x(0) = xo

then e(?) satisfies the following non-autonomous differential equation

e = f (x* + e, t) - f(x*. t) = g(e, t) (3.8)

with initial condition e(0) = 8x0. Since g(0, t) = 0, the new dynamic system, with e asstate and g in place of f, has an equilibrium point at the origin of the state space.Therefore, instead of studying the deviation of x(0 from x (f) for the original system,we may simply study the stability of the perturbation dynamics (3.8) with respect tothe equilibrium point 0. Note, however, that the perturbation dynamics is non-autonomous, due to the presence of the nominal trajectory x*(r) on the right-hand side.Each particular nominal motion of an autonomous system corresponds to anequivalent non-autonomous system, whose study requires the non-autonomous systemanalysis techniques to be presented in chapter 4.

Let us now illustrate this important transformation on a specific system.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (61)

Sect. 3.2 Concepts of Stability 47

Example 3.2: Consider the autonomous mass-spring system

3 = 0

which contains a nonlinear term reflecting the hardening effect of the spring. Let us study the

stability of the motion x*(t) which starts from initial position x0.

Assume that we slightly perturb the initial position to be x(0) = xo + 5x0. The resulting system

trajectory is denoted as x(t). Proceeding as before, the equivalent differential equation governing

the motion error e is

me + k\ e + k2 I e3 + 3e2x*(t) + 3ex*2(t) ] = 0

Clearly, this is a non-autonomous system. •

Of course, one can also show that for non-autonomous nonlinear systems, thestability problem around a nominal motion can also be transformed as a stabilityproblem around the origin for an equivalent non-autonomous system.

Finally, note that if the original system is autonomous and linear, in the form(3.4), then the equivalent system is still autonomous, since it can be written

e = Ae

3.2 Concepts of Stability

In the beginning of this chapter, we introduced the intuitive notion of stability as akind of well-behavedness around a desired operating point. However, since nonlinearsystems may have much more complex and exotic behavior than linear systems, themere notion of stability is not enough to describe the essential features of their motion.A number of more refined stability concepts, such as asymptotic stability, exponentialstability and global asymptotic stability, are needed. In this section, we define thesestability concepts formally, for autonomous systems, and explain their practicalmeanings.

A few simplifying notations are defined at this point. Let B^ denote thespherical region (or ball) defined by || x ]| < R in state-space, and S/j the sphere itself,defined by || x || = R.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (62)

48 Fundamentals of Lyapunov Theory Chap. 3

STABILITY AND INSTABILITY

Let us first introduce the basic concepts of stability and instability.

Definition 3.3 The equilibrium state x = 0 is said to be stable if, for any R>0, thereexists r>0, such that if ||x(0)|| < r, then ||x(f)|| <R for all t>0 . Otherwise, theequilibrium point is unstable.

Essentially, stability (also called stability in the sense of Lyapunov, or Lyapunovstability) means that the system trajectory can be kept arbitrarily close to the origin bystarting sufficiently close to it. More formally, the definition states that the origin isstable, if, given that we do not want the state trajectory x(f) to get out of a ball ofarbitrarily specified radius B^ , a value r(R) can be found such that starting the statefrom within the ball B,. at time 0 guarantees that the state will stay within the ball B^thereafter. The geometrical implication of stability is indicated by curve 2 in Figure3.3. Chapter 2 provides examples of stable equilibrium points in the case of second-order systems, such as the origin for the mass-spring system of Example 2.1, or stablenodes or foci in the local linearization of a nonlinear system.

Throughout the book, we shall use the standard mathematical abbreviationsymbols:

V to mean "for any"3 for "there exists"e for "in the set"=> for "implies that"

Of course, we shall say interchangeably that A implies B, or that A is a sufficientcondition of B, or that B is a necessary condition of A. If A => B and B => A ,then A and B are equivalent, which we shall denote by A <=> B .

Using these symbols, Definition 3.3 can be written

VR>0,3r>0, || x(0) \\<r => V t > 0 , || x(f) || < R

or, equivalently

V t f > 0 , 3 r > 0 , x(0) e Br => V t > 0 , x(t) e BR

Conversely, an equilibrium point is unstable if there exists at least one ball B^,such that for every r>0, no matter how small, it is always possible for the systemtrajectory to start somewhere within the ball Br and eventually leave the ball BR

(Figure 3.3). Unstable nodes or saddle points in second-order systems are examples ofunstable equilibria. Instability of an equilibrium point is typically undesirable, because

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (63)

Sect. 3.2 Concepts of Stability 49

it often leads the system into limit cycles or results in damage to the involvedmechanical or electrical components.

curve 1 - asymptotically stable

curve 2 - marginally stable

curve 3 - unstable

Figure 3.3 : Concepts of stability

It is important to point out the qualitative difference between instability and theintuitive notion of "blowing up" (all trajectories close to origin move further andfurther away to infinity). In linear systems, instability is equivalent to blowing up,because unstable poles always lead to exponential growth of the system states.However, for nonlinear systems, blowing up is only one way of instability. Thefollowing example illustrates this point.

Example 3.3: Instability of the Van der Pol Oscillator

The Van der Pol oscillator of Example 2.6 is described by

One easily shows that the system has an equilibrium point at the origin.

As pointed out in section 2.2 and seen in the phase portrait of Figure 2.8, system trajectories

starting from any non-zero initial states all asymptotically approach a limit cycle. This implies

that, if we choose R in Definition 3.3 to be small enough for the circle of radius R to fall

completely within the closed-curve of the limit cycle, then system trajectories starting near the

origin will eventually get out of this circle (Figure 3.4). This implies instability of the origin.

Thus, even though the state of the system does remain around the equilibrium point in a

certain sense, it cannot stay arbitrarily close to it. This is the fundamental distinction between

stability and instability. d

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (64)

50 Fundamentals of Lyapunov Theory Chap. 3

trajectories

limitcycle

Figure 3.4 : Unstable origin of the Van der Pol Oscillator

ASYMPTOTIC STABILITY AND EXPONENTIAL STABILITY

In many engineering applications, Lyapunov stability is not enough. For example,when a satellite's attitude is disturbed from its nominal position, we not only want thesatellite to maintain its attitude in a range determined by the magnitude of thedisturbance, i.e., Lyapunov stability, but also require that the attitude gradually goback to its original value. This type of engineering requirement is captured by theconcept of asymptotic stability.

Definition 3.4 An equilibrium point 0 is asymptotically stable if it is stable, and if inaddition there exists some r > 0 such that || x(0) || < r implies that \{t) —> 0 as t —> °°.

Asymptotic stability means that the equilibrium is stable, and that in addition,states started close to 0 actually converge to 0 as time t goes to infinity. Figure 3.3shows that system trajectories starting from within the ball B,. converge to the origin.The ball Br is called a domain of attraction of the equilibrium point (while the domainof attraction of the equilibrium point refers to the largest such region, i.e., to the set ofall points such that trajectories initiated at these points eventually converge to theorigin). An equilibrium point which is Lyapunov stable but not asymptotically stableis called marginally stable.

One may question the need for the explicit stability requirement in thedefinition above, in view of the second condition of state convergence to the origin.However, it it easy to build counter-examples that show that state convergence doesnot necessarily imply stability. For instance, a simple system studied by Vinograd hastrajectories of the form shown in Figure 3.5. All the trajectories starting from non-zero

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (65)

Sect. 3.2 Concepts of Stability 51

initial points within the unit disk first reach the curve C before converging to theorigin. Thus, the origin is unstable in the sense of Lyapunov, despite the stateconvergence. Calling such a system unstable is quite reasonable, since a curve such asC may be outside the region where the model is valid - for instance, the subsonic andsupersonic dynamics of a high-performance aircraft are radically different, while, withthe problem under study using subsonic dynamic models, C could be in the supersonicrange.

Figure 3.5 : State convergence does not imply stability

In many engineering applications, it is still not sufficient to know that a systemwill converge to the equilibrium point after infinite time. There is a need to estimatehow fast the system trajectory approaches 0. The concept of exponential stability canbe used for this purpose.

Definition 3.5 An equilibrium point 0 is exponentially stable if there exist twostrictly positive numbers a and X such that

Vf>0, ||x(?)|| < a| |x(O)| |e-^' 0-9)

in some ball Br around the origin.

In words, (3.9) means that the state vector of an exponentially stable systemconverges to the origin faster than an exponential function. The positive number X isoften called the rate of exponential convergence. For instance, the system

x = - ( 1 + sin2*) x

is exponentially convergent to x = 0 with a rate X = 1 . Indeed, its solution is

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (66)

52 Fundamentals ofLyapunov Theory Chap. 3

x(t) = x(0) exp(- ['[ 1 + sin2(jc(x))] dx)

and therefore

\x(0)\e-<

Note that exponential stability implies asymptotic stability. But asymptoticstability does not guarantee exponential stability, as can be seen from the system

x = -x2, 40) =1 (3.10)

whose solution is x = 1/(1 + t), a function slower than any exponential function e~^f

(with X > 0).

The definition of exponential convergence provides an explicit bound on thestate at any time, as seen in (3.9). By writing the positive constant a as a = e^xo , it iseasy to see that, after a time of xo + (l/X), the magnitude of the state vector decreasesto less than 35% ( ~ e~ ' ) of its original value, similarly to the notion of time-constantin a linear system. After \0 + (3/X.), the state magnitude ||x(r)|| will be less than5% ( = e - 3 )of

LOCAL AND GLOBAL STABILITY

The above definitions are formulated to characterize the local behavior of systems,i.e., how the state evolves after starting near the equilibrium point. Local propertiestell little about how the system will behave when the initial state is some distanceaway from the equilibrium, as seen for the nonlinear system in Example 1.1. Globalconcepts are required for this purpose.

Definition 3.6 If asymptotic (or exponential) stability holds for any initial states, theequilibrium point is said to be asymptotically (or exponentially) stable in the large. Itis also called globally asymptotically (or exponentially) stable.

For instance, in Example 1.2 the linearized system is globally asymptoticallystable, but the original system is not. The simple system in (3.10) is also globallyasymptotically stable, as can be seen from its solutions.

Linear time-invariant systems are either asymptotically stable, or marginallystable, or unstable, as can be be seen from the modal decomposition of linear systemsolutions; linear asymptotic stability is always global and exponential, and linearinstability always implies exponential blow-up. This explains why the refined notionsof stability introduced here were not previously encountered in the study of linearsystems. They are explicitly needed only for nonlinear systems.

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (67)

Sect. 3.3 Linearization and Local Stability 53

3.3 Linearization and Local Stability

Lyapunov's linearization method is concerned with the local stability of a nonlinearsystem. It is a formalization of the intuition that a nonlinear system should behavesimilarly to its linearized approximation for small range motions. Because all physicalsystems are inherently nonlinear, Lyapunov's linearization method serves as thefundamental justification of using linear control techniques in practice, i.e., shows thatstable design by linear control guarantees the stability of the original physical systemlocally.

Consider the autonomous system in (3.2), and assume that f(x) is continuouslydifferentiable. Then the system dynamics can be written as

where fh 01 stands for higher-order terms in x. Note that the above Taylor expansionstarts directly with the first-order term, due to the fact that f(0) = 0 , since 0 is anequilibrium point. Let us use the constant matrix A to denote the Jacobian matrix of fwith respect to x at x = 0 (an nx n matrix of elements 3/j- / dxj)

Then, the system

x = A x (3.12)

is called the linearization (or linear approximation) of the original nonlinear system atthe equilibrium point 0.

Note that, similarly, starting with a non-autonomous nonlinear system with a control input u

x = f(x, u)

such that f(0, 0) = 0 , we can write

x = ( — ) x + ( — ) u + fh o t (x, u)V9x/(x=0,u=0) V3u/ ( x =0,u = 0)

where fh n t stands for higher-order terms in x and u. Letting A denote the Jacobian matrix of f withrespect to x at (x = 0, u = 0) , and B denote the Jacobian matrix of f with respect to u at the samepoint (annxm matrix of elements dft I duj , where m is the number of inputs)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (68)

54 Fundamentals of Lyapunov Theory Chap. 3

'(x=O,u=O) v d u / ( x = o , u = O )

the system

x = A x + B u

is the linearization (or linear approximation) of the original nonlinear system at (x = 0, u = 0 ) .

Furthermore, the choice of a control law of the form u = u(x) (with u(0) = 0 ) transforms the

original non-autonomous system into an autonomous closed-loop system, having x = 0 as an

equilibrium point. Linearly approximating the control law as

^ x = Gx

the closed-loop dynamics can be linearly approximated as

x = f(x,u(x)) « (A + B G ) x

Of course, the same linear approximation can be obtained by directly considering the autonomous

closed-loop system

x = f(x, u(x)) = f,(x)

and linearizing the function f | with respect to x, at its equilibrium point x = 0.

In practice, finding a system's linearization is often most easily done simply byneglecting any term of order higher than 1 in the dynamics, as we now illustrate.

Example 3.4: Consider the system

x\ ~ X22 + x\ c o s ; t 2

x2 = x2 + (x{ + l)xl+xl sin x2

Its linearized approximation about x = 0 is

i | = 0 + Xj • 1 = Jtj

X2 ~ X2 + 0 + x\ +x\ X2 ~ X2 + x\

The linearized system can thus be written

1 0x = x

1 1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (69)

Sect. 3.3 Linearization and Local Stability 55

A similar procedure can be applied for a controlled system. Consider the system

x + 4xs + (x2 + 1) w = 0

The system can be linearly approximated about x = 0 as

i.e., the linearized system can be written

x = -u

Assume that the control law for the original nonlinear system has been selected to be

u = sinx + x 3 + x c o s 2 x

then the linearized closed-loop dynamics is

The following result makes precise the relationship between the stability of thelinear system (3.12) and that of the original nonlinear system (3.2).

Theorem 3.1 (Lyapunov's linearization method)

• // the linearized system is strictly stable (i.e, if all eigenvalues of A arestrictly in the left-half complex plane), then the equilibrium point isasymptotically stable (for the actual nonlinear system).

• If the linearized system is unstable (i.e, if at least one eigenvalue of A isstrictly in the right-half complex plane), then the equilibrium point is unstable(for the nonlinear system).

• If the linearized system is marginally stable (i.e, all eigenvalues of A are inthe left-half complex plane, but at least one of them is on the /co axis), thenone cannot conclude anything from the linear approximation (the equilibriumpoint may be stable, asymptotically stable, or unstable for the nonlinearsystem).

While the proof of this theorem (which is actually based on Lyapunov's directmethod, see Exercise 3.12) shall not be detailed, let us remark that its results areintuitive. A summary of the theorem is that it is true by continuity. If the linearizedsystem is strictly stable, or strictly unstable, then, since the approximation is valid "nottoo far" from the equilibrium, the nonlinear system itself is locally stable, or locallyunstable. However, if the linearized system is marginally stable, the higher-orderterms in (3.11) can have a decisive effect on whether the nonlinear system is stable or

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (70)

56 Fundamentals of Lyapunov Theory Chap. 3

unstable. As we shall see in the next section, simple nonlinear systems may beglobally asymptotically stable while their linear approximations are only marginallystable: one simply cannot infer any stability property of a nonlinear system from itsmarginally stable linear approximation.

Example 3.5: As expected, it can be shown easily that the equilibrium points (8 = n [2n] ,9 = 0)of the pendulum of Example 3.1 are unstable. Consider for instance the equilibrium point(9 = 7t, 6 = 0). Since, in a neighborhood of 6 = 7t, we can write

sin 6 = sin 7t + cos n (6 — n) + h.o.t. = (7t — 8) + h.o.t.

thus, letting 8 = 6 - TC , the system's linearization about the equilibrium point (9 = n , 8 = 0) is

H - b s _ £ § = o

I

MR2

Hence the linear approximation is unstable, and therefore so is the nonlinear system at this

equilibrium point. LJ

Example 3.6: Consider the first order system

The origin 0 is one of the two equilibrium points of this system. The linearization of this system

around the origin is

The application of Lyapunov's linearization method indicates the following stability properties of

the nonlinear system

• a < 0 : asymptotically stable;

• a > 0 : unstable;

• o = 0 : cannot tell from linearization.

In the third case,, the nonlinear system is

x = bx5

The linearization method fails while, as we shall see, the direct method to be described can easily

solve this problem. CD

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (71)

Sect. 3.4 Lyapunov's Direct Method 57

Lyapunov's linearization theorem shows that linear control design is a matter ofconsistency: one must design a controller such that the system remain in its "linearrange". It also stresses major limitations of linear design: how large is the linearrange? What is the extent of stability (how large is r in Definition 3.3) ? Thesequestions motivate a deeper approach to the nonlinear control problem, Lyapunov'sdirect method.

3.4 Lyapunov's Direct Method

The basic philosophy of Lyapunov's direct method is the mathematical extension of afundamental physical observation: if the total energy of a mechanical (or electrical)system is continuously dissipated, then the system, whether linear or nonlinear, musteventually settle down to an equilibrium point. Thus, we may conclude the stability ofa system by examining the variation of a single scalar function.

Specifically, let us consider the nonlinear mass-damper-spring system in Figure3.6, whose dynamic equation is

= 0 (3.13)

with bx\x\ representing nonlinear dissipation or damping, and (kox +representing a nonlinear spring term. Assume that the mass is pulled away from thenatural length of the spring by a large distance, and then released. Will the resultingmotion be stable? It is very difficult to answer this question using the definitions ofstability, because the general solution of this nonlinear equation is unavailable. Thelinearization method cannot be used either because the motion starts outside the linearrange (and in any case the system's linear approximation is only marginally stable).However, examination of the system energy can tell us a lot about the motion pattern.

nonlinearspring and

p damper

x Figure 3.6 : A nonlinear mass-damper-spring system

The total mechanical energy of the system is the sum of its kinetic energy andits potential energy

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (72)

58 Fundamentals ofLyapunov Theory Chap. 3

V(x) =- mx2 + \"(kox + k{x3) dx = -m'x2 + -k0 x2+]-kxx

A (3.14)Z J o L 2. <\

Comparing the definitions of stability and mechanical energy, one can easily see somerelations between the mechanical energy and the stability concepfs described earlier:

• zero energy corresponds to the equilibrium point (x = 0, x = 0)• asymptotic stability implies the convergence of mechanical energy to zero• instability is related to the growth of mechanical energy

These relations indicate that the value of a scalar quantity, the mechanical energy,indirectly reflects the magnitude of the state vector; and furthermore, that the stabilityproperties of the system can be characterized by the variation of the mechanicalenergy of the system.

The rate of energy variation during the system's motion is obtained easily bydifferentiating the first equality in (3.14) and using (3.13)

V(x) = mxx + (kox + kx x3) x = x (-6x1x1) = -fclxl3 (3.15)

Equation (3.15) implies that the energy of the system, starting from some initial value,is continuously dissipated by the damper until the mass settles down, i.e., until x = 0.Physically, it is easy to see that the mass must finally settle down at the natural lengthof the spring, because it is subjected to a non-zero spring force at any position otherthan the natural length.

The direct method of Lyapunov is based on a generalization of the concepts inthe above mass-spring-damper system to more complex systems. Faced with a set ofnonlinear differential equations, the basic procedure of Lyapunov's direct method is togenerate a scalar "energy-like" function for the dynamic system, and examine the timevariation of that scalar function. In this way, conclusions may be drawn on thestability of the set of differential equations without using the difficult stabilitydefinitions or requiring explicit knowledge of solutions.

3.4.1 Positive Definite Functions and Lyapunov Functions

The energy function in (3.14) has two properties. The first is a property of the functionitself: it is strictly positive unless both state variables x and x are zero. The secondproperty is a property associated with the dynamics (3.13): the function ismonotonically decreasing when the variables x and x vary according to (3.13). InLyapunov's direct method, the first property is formalized by the notion of positivedefinite functions, and the second is formalized by the so-called Lyapunov functions.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (73)

Sect. 3.4 Lyapunov's Direct Method 59

Let us discuss positive definite functions first.

Definition 3.7 A scalar continuous function V(x) is said to be locally positivedefinite ifV(0) = 0 and, in a ball BR

x jt 0 => V(x) > 0

//V(0) = 0 and the above property holds over the whole state space, then V{\) is saidto be globally positive definite.

For instance, the function

V(x) = ~MR2x22 + MRgil-cosx^

which is the mechanical energy of the pendulum of Example 3.1, is locally positivedefinite. The mechanical energy (3.14) of the nonlinear mass-damper-spring system isglobally positive definite. Note that, for that system, the kinetic energy (1/2) m'x2 isnot positive definite by itself, because it can equal zero for non-zero values of X

The above definition implies that the function V has a unique minimum at theorigin 0. Actually, given any function having a unique minimum in a certain ball, wecan construct a locally positive definite function simply by adding a constant to thatfunction. For example, the function V(x) = x2 + x2

2 - 1 is a lower bounded functionwith a unique minimum at the origin, and the addition of the constant 1 to it makes it apositive definite function. Of course, the function shifted by a constant has the sametime-derivative as the original function.

Let us describe the geometrical meaning of locally positive definite functions.Consider a positive definite function V(x) of two state variables Xj and x2. Plotted in a3-dimensional space, V(x) typically corresponds to a surface looking like an upwardcup (Figure 3.7). The lowest point of the cup is located at the origin.

A second geometrical representation can be made as follows. Taking Xj and x2

as Cartesian coordinates, the level curves V(x\,xj) = Va typically represent a set ofovals surrounding the origin, with each oval corresponding to a positive value of Va .These ovals, often called contour curves, may be thought as the sections of the cup byhorizontal planes, projected on the (xj, x2) plane (Figure 3.8). Note that the contourcurves do not intersect, because V(xt, x2) is uniquely defined given (xj , x2).

A few related concepts can be defined similarly, in a local or global sense, i.e., afunction V(x) is negative definite if — V(x) is positive definite; V(x) is positivesemi-definite if V(0) = 0 and V(x) > 0 for x * 0; V(x) is negative semi-definite if - V(x)is positive semi-definite. The prefix "semi" is used to reflect the possibility of V being

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (74)

60 Fundamentals of Lyapunov Theory

v=vl

V2> \

Chap. 3

Figure 3.7 : Typical shape of a positive definite function V(JCJ,

equal to zero for x 0. These concepts can be given geometrical meanings similar tothe ones given for positive definite functions.

With x denoting the state of the system (3.2), a scalar function V(x) actuallyrepresents an implicit function of time t. Assuming that V(x) is differentiable, itsderivative with respect to time can be found by the chain rule,

,-, dV(x) 3 V . 9 y t , ,V = ^ = — x = — f(x)3 3dt 3 x 3 x

V=

Figure 3.8 : Interpreting positive definite functions using contour curves

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (75)

Sect. 3.4 Lyapunov's Direct Method 61

We see that, because x is required to satisfy the autonomous state equations (3.2),V only depends on x. It is often referred to as "the derivative of V along the systemtrajectory" - in particular, V = 0 at an equilibrium point. For the system (3.13), V x)is computed in (3.15) and found to be negative. Functions such as V in that exampleare given a special name because of their importance in Lyapunov's direct method.

Definition 3.8 If, in a ball BR , the function V(x) is positive definite and hascontinuous partial derivatives, and if its time derivative along any state trajectory ofsystem (3.2) is negative semi-definite, i.e.,

V(x) < 0

then V(x) is said to be a Lyapunov function for the system (3.2).

Figure 3.9 : Illustrating Definition 3.8 for n = 2

A Lyapunov function can be given simple geometrical interpretations. InFigure 3.9, the point denoting the value of V(xj,x2) is seen to always point down abowl. In Figure 3.10, the state point is seen to move across contour curvescorresponding to lower and lower values of V.

3.4.2 Equilibrium Point Theorems

The relations between Lyapunov functions and the stability of systems are madeprecise in a number of theorems in Lyapunov's direct method. Such theorems usuallyhave local and global versions. The local versions are concerned with stabilityproperties in the neighborhood of equilibrium point and usually involve a locallypositive definite function.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (76)

62 Fundamentals of Lyapunov Theory Chap. 3

V = V

V = V

Figure 3.10 : Illustrating Definition 3.8 for n = 2 using contour curves

LYAPUNOV THEOREM FOR LOCAL STABILITY

Theorem 3.2 (Local Stability) //, in a ball hR , there exists a scalar function V(x)with continuous first partial derivatives such that

• V(x) is positive definite (locally in BR )

• V(x) is negative semi-definite (locally in BR )

then the equilibrium point 0 is stable. If, actually, the derivative V(x) is locallynegative definite in B^ , then the stability is asymptotic.

The proof of this fundamental result is conceptually simple, and is typical ofmany proofs in Lyapunov theory.

Proof: Let us derive the result using the geometric interpretation of a Lyapunov function, as

illustrated in Figure 3.9 in the case n = 2. To show stability, we must show that given any strictly

positive number R, there exists a (smaller) strictly positive number r such that any trajectory

starting inside the ball B r remains inside the ball BR for all future time. Let m be the minimum of

V on the sphere SR . Since V is continuous and positive definite, m exists and is strictly positive.

Furthermore, since K(0) = 0, there exists a ball B r around the origin such that V(x) < m for any x

inside the ball (Figure 3.1 la). Consider now a trajectory whose initial point x(0) is within the ball

B r . Since V is non-increasing along system trajectories, V remains strictly smaller than m, and

therefore the trajectory cannot possibly cross the outside sphere S# . Thus, any trajectory starting

inside the ball B r remains inside the ball B^ , and therefore Lyapunov stability is guaranteed.

Let us now assume that V is negative definite, and show asymptotic stability, by

contradiction. Consider a trajectory starting in some ball B r as constructed above (e.g., the ball B r

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (77)

Sect. 3.4 Lyapunov's Direct Method 63

(a) (b)

Figure 3 .11 : Illustrating the proof of Theorem 3.2 for n = 2

corresponding to R = Ro). Then the trajectory will remain in the ball B s for all future time. Since

V is lower bounded and decreases continually, V tends towards a limit L, such that

V r > 0 , V(x(t))>L. Assume that this limit is not zero, i.e., that L > 0 . Then, since V is

continuous and V(0) = 0, there exists a ball B r that the system trajectory never enters (Figure

3.1 lb). But then, since - V is also continuous and positive definite, and since BR is bounded,

- V must remain larger than some strictly positive number L ; . This is a contradiction, because it

would imply that V(i) decreases from its initial value Vg to a value strictly smaller than L, in a

finite time smaller than [Vg - L)/L\. Hence, all trajectories starting in B r asymptotically converge

to the origin. Q

In applying the above theorem for analysis of a nonlinear system, one goes

through the two steps of choosing a positive definite function, and then determining its

derivative along the path of the nonlinear systems. The following example illustrates

this procedure.

Example 3.7: Local Stability

A simple pendulum with viscous damping is described by

e + e + sin e = o

Consider the following scalar function

V(x) = ( l -

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (78)

64 Fundamentals of Lyapunov Theory Chap. 3

One easily verifies that this function is locally positive definite. As a matter of fact, this function

represents the total energy of the pendulum, composed of the sum of the potential energy and the

kinetic energy. Its time-derivative is easily found to be

v(x) = esine + ee = - e 2 s o

Therefore, by invoking the above theorem, one concludes that the origin is a stable equilibrium

point. In fact, using physical insight, one easily sees the reason why V(x) < 0, namely that the

damping term absorbs energy. Actually, V is precisely the power dissipated in the pendulum.

However, with this Lyapunov function, one cannot draw conclusions on the asymptotic stability

of the system, because V x) is only negative semi-definite. Lj

The following example illustrates the asymptotic stability result.

Example 3.8: Asymptotic stability

Let us study the stability of the nonlinear system defined by

Jcj = j t | ( x j 2 + x22— 2) — Ax\X2

2

x2 = Ax^Xj + x2 (X[2 + x2 - 2)

around its equilibrium point at the origin. Given the positive definite function

V{x\,x2)=xl2+x2

2

its derivative V along any system trajectory is

Thus, V is locally negative definite in the 2-dimensional ball B2, i.e., in the region defined by

jf]2 + x22 < 2. Therefore, the above theorem indicates that the origin is asymptotically stable. Ll

LYAPUNOV THEOREM FOR GLOBAL STABILITY

The above theorem applies to the local analysis of stability. In order to assert globalasymptotic stability of a system, one might naturally expect that the ball B^ in theabove local theorem has to be expanded to be the whole state-space. This is indeednecessary, but it is not enough. An additional condition on the function V has to besatisfied: V(x) must be radially unbounded, by which we mean that V(x) —> °° as||x|| —> °° (in other words, as x tends to infinity in any direction). We then obtain thefollowing powerful result:

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (79)

Sect. 3.4 Lyapuno v 's Direct Method 65

Theorem 3.3 (Global Stability) Assume that there exists a scalar function V of thestate x, with continuous first order derivatives such that

• V(x) is positive definite

• V(x) is negative definite

• V(x) -^oo as \\x\\ -> °°

then the equilibrium at the origin is globally asymptotically stable.

Proof: The proof is the same as in the local case, by noticing that the radial unboundedness of V,

combined with the negative defmiteness of V, implies that, given any initial condition xo, the

trajectories remain in the bounded region defined by V(x) < V(\g). O

The reason for the radial unboundedness condition is to assure that the contourcurves (or contour surfaces in the case of higher order systems) V(x) = Va correspondto closed curves. If the curves are not closed, it is possible for the state trajectories todrift away from the equilibrium point, even though the state keeps going throughcontours corresponding to smaller and smaller Va

>s- For example, for the positivedefinite function V = [xj2/(l + Xj2)] + x2

2, the curves V(x) = Va for Va > 1 are opencurves. Figure 3.12 shows the divergence of the state while moving toward lower andlower "energy" curves. Exercise 3.4 further illustrates this point on a specific system.

V(x) = K.

V(x) = K,V(x) = V.

V > V > V1 2 3

Figure 3.12 : Motivation of the radial unboundedness condition

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (80)

66 Fundamentals of Lyapunov Theory

Example 3.9: A class of first-order systems

Consider the nonlinear system

x + c(x) = 0

where c is any continuous function of the same sign as its scalar argument x, i.e.,

xc(x) > 0 for

Chap. 3

Intuitively, this condition indicates that - c(x) "pushes" the system back towards its rest position

x = 0, but is otherwise arbitrary. Since c is continuous, it also implies that c(0) = 0 (Figure 3.13).

Consider as the Lyapunov function candidate the square of the distance to the origin

The function V is radially unbounded, since it tends to infinity as \x\ —» ° ° . Its derivative is

Thus V < 0 as long as x * 0, so that x = 0 is a globally asymptotically stable equilibrium point.

c(x)

Figure 3.13 : The function c(x)

For instance, the system

is globally asymptotically convergent to x = 0, since for x * 0 , sin2 x < (sin JC| -c (JC|. Similarly, the

system

is globally asymptotically convergent to x = 0. Notice that while this system's linear

approximation ( x ~ 0 ) is inconclusive, even about local stability, the actual nonlinear system

enjoys a strong stability property (global asymptotic stability). O

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (81)

Sect. 3.4 Lyapunov's Direct Method 67

Example 3.10: Consider the system

x\ =X2~X\(X\2 + X22)

The origin of the state-space is an equilibrium point for this system. Let V be the positive definite

function

V(x) = x2 + x2

The derivative of V along any system trajectory is

V(x) = 2x, xl + 2x2i2 = -2(xx2 + x2

2)2

which is negative definite. Therefore, the origin is a globally asymptotically stable equilibrium

point. Note that the globalness of this stability result also implies that the origin is the only

equilibrium point of the system. f~1

REMARKS

Many Lyapunov functions may exist for the same system. For instance, if V is aLyapunov function for a given system, so is

Vx = pVa

where p is any strictly positive constant and a is any scalar (not necessarily an integer)larger than 1. Indeed, the positive-definiteness of V implies that of Vj , the positive-definiteness (or positive semi-definiteness) of - V implies that of —V\, and (theradial unboundedness of V (if applicable) implies that of Vx .

More importantly, for a given system, specific choices of Lyapunov functionsmay yield more precise results than others. Consider again the pendulum of Example3.7. The function

V(x) = - 02 + 1 (6 + 6)2 + 2( 1 - cosG)

is also a Lyapunov function for the system, because locally

i 2 < 0

However, it is interesting to note that V is actually locally negative definite, andtherefore, this modified choice of V, without obvious physica! meaning, allows theasymptotic stability of the pendulum to be shown.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (82)

68 Fundamentals of Lyapunov Theory Chap. 3

Along the same lines, it is important to realize that the theorems in Lyapunovanalysis are all sufficiency theorems. If for a particular choice of Lyapunov functioncandidate V, the conditions on V are not met, one cannot draw any conclusions on thestability or instability of the system - the only conclusion one should draw is that adifferent Lyapunov function candidate should be tried.

3.4.3 Invariant Set Theorems

Asymptotic stability of a control system is usually a very important property to bedetermined. However, the equilibrium point theorems just described are often difficultto apply in order to assert this property. The reason is that it often happens that V, thederivative of the Lyapunov function candidate, is only negative semi-definite, as seenin (3.15). In this kind of situation, fortunately, it is still possible to draw conclusionson asymptotic stability, with the help of the powerful invariant set theorems, attributedto La Salle. This section presents the local and global versions of the invariant settheorems.

The central concept in these theorems is that of invariant set, a generalization ofthe concept of equilibrium point.

Definition 3.9 A set G is an invariant set for a dynamic system if every systemtrajectory which starts from a point in G remains in G for all future time.

For instance, any equilibrium point is an invariant set. The domain of attraction of anequilibrium point is also an invariant set. A trivial invariant set is the whole state-space. For an autonomous system, any of the trajectories in state-space is an invariantset. Since limit cycles are special cases of system trajectories (closed curves in thephase plane), they are also invariant sets.

Besides often yielding conclusions on asymptotic stability when V, thederivative of the Lyapunov function candidate, is only negative semi-definite, theinvariant set theorems also allow us to extend the concept of Lyapunov function so asto describe convergence to dynamic behaviors more general than equilibrium, e.g.,convergence to a limit cycle.

Similarly to our earlier discussion of Lyapunov's direct method, we first discussthe local version of the invariant set theorems, and then the global version.

LOCAL INVARIANT SET THEOREM

The invariant set theorems reflect the intuition that the decrease of a Lyapunovfunction V has to gradually vanish (i.e., V has to converge to zero) because V is lower

I

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (83)

Sect. 3.4 Lyapunov's Direct Method 69

bounded. A precise statement of this result is as follows.

Theorem 3.4 (Local Invariant Set Theorem) Consider an autonomous system ofthe form (3.2), with f continuous, and let V(x) be a scalar function with continuousfirst partial derivatives. Assume that

• for some I > 0, the region Qj defined by V(x) < I is bounded

• V(x) < 0 for all x in Q.t

Let R be the set of all points within Q[ where V(x) = 0, and M be the largest

invariant set in R. Then, every solution x(t) originating in £2 tends to M as t —» °°.

In the above theorem, the word "largest" is understood in the sense of set theory, i.e.,M is the union of all invariant sets (e.g., equilibrium points or limit cycles) within R.In particular, if the set R is itself invariant (i.e., if once V= 0, then V = 0 for all futuretime), then M = R. Also note that V, although often still referred to as a Lyapunovfunction, is not required to be positive definite.

The geometrical meaning of the theorem is illustrated in Figure 3.14, where atrajectory starting from within the bounded region Q ; is seen to converge to the largestinvariant set M. Note that the set R is not necessarily connected, nor is the set M.

V = /

M

Figure 3.14 : Convergence to the largest invariant set M

The theorem can be proven in two steps, by first showing that V goes to zero,and then showing that the state converges to the largest invariant set within the setdefined by V=0. We shall simply give a sketch of the proof, since the detailed proofof the second part involves a number of concepts in topology and real analysis whichare not prerequisites of this text.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (84)

70 Fundamentals of Lyapunov Theory Chap. 3

Proof: The first part of the proof involves showing that V —> 0 for any trajectory starting from a

point in Qy, using a result in functional analysis known as Barbalat's lemma, which we shall

detail in section 4.3.

Specifically, consider a trajectory starting from an arbitrary point x0 in £ij. The trajectory

must stay in Q ; all the time, because V < 0 implies that V[x(t)] < V[x(0)] < / for all t>0. In

addition, because V(x) is continuous in x (since it is differentiable with respect to x) over the

bounded region Q , it is lower bounded in that region; therefore, since we just noticed that the

trajectory remains in Q ; , V[x(r)l remains lower bounded for all ( > 0. Furthermore, the facts that

f is continuous, V has continuous partial derivatives, and the region Q.j is bounded, imply that V is

uniformly continuous. Therefore, V[x(t)] satisfies the three conditions (V lower bounded; V < 0;

V uniformly continuous) of Barbalat's lemma. As a result, V[x(()] - • 0, which implies that all

system trajectories starting from within Qy converge to the set R.

The second part of the proof [see, e.g., Hahn, 1968] involves showing that the trajectories

cannot converge to just anywhere in the set R: they must converge to the largest invariant set M

within R. This can be proven by showing that any bounded trajectory of an autonomous system

converges to an invariant set (the so-called positive limit set of the trajectory), and then simply

noticing that this set is a subset of the largest invariant set M. D

Note that the asymptotic stability result in the local Lyapunov theorem can beviewed a special case of the above invariant set theorem, where the set M consistsonly of the origin.

Let us now illustrate applications of the invariant set theorem using someexamples. The first example shows how to conclude asymptotic stability for problemswhich elude the local Lyapunov theorem. The second example shows how todetermine a domain of attraction, an issue which was not specifically addressedbefore. The third example shows the convergence of system trajectories to a limitcycle.

Example 3.11: Asymptotic stability of the mass-damper-spring system

For the system (3.13), one can only draw conclusion of marginal stability using the energy

function (3.14) in the local equilibrium point theorem, because V is only negative semi-definite

according to (3.15). Using the invariant set theorem, however, we can show that the system is

actually asymptotically stable. To do this, we only have to show that the set M contains only one

point.

The set R is defined by x = 0, i.e., the collection of states with zero velocity, or the whole

horizontal axis in the phase plane (x, x). Let us show that the largest invariant set M in this set R

contains only the origin. Assume that M contains a point with a nonzero position *,, then, the

acceleration at that point is ji" = - (kolm)x- (k\/m)x^ * 0 . This implies that the trajectory will

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (85)

Sect. 3.4 Lyapunov's Direct Method 71

immediately move out of the set R and thus also out of the set M , a contradict ion to the

definition. CD

Example 3.12: D o m a i n of Attract ion

Consider again the system in Example 3.8. For / = 2, the region £ i 2 > defined by

V(x) = X | 2 + x 2 < 2 , is bounded. The set R is s imply the origin 0, which is an invariant set

(since it is an equil ibrium point). All the condit ions of the local invariant set theorem are satisfied

and, therefore, any trajectory starting within the circle converges to the origin. Thus, a domain of

attraction is explicitly determined by the invariant set theorem. CD

Example 3 .13: Attractive Limit Cycle

Consider the system

xx = x2 - x^ [x[A + 2x2

2- 1 0 ]

x2 = - x , 3 - 3 x 25 [ x , 4 + 2 x 2

2 - 10]

Notice first that the set defined by xt4 + 2x 2

2 = 10 is invariant, since

- ( x , 4 + 2 x 22 - 10) = - ( 4 x , 1 0 + 12x2

6)(x,4 + 2 x 22 - 10)

which is zero on the set. The motion on this invariant set is described (equivalently) by either of

the equations

x, = x2

x2 = - x , 3

Therefore, we see that the invariant set actually represents a limit cycle, along which the state

vector moves clockwise.

Is this limit cycle actually attractive? Let us define as a Lyapunov function candidate

l /=(x, 4 + 2 x 22 - I0)2

which represents a measure of the "distance" to the limit cycle. For any arbitrary positive number

/, the region Q/, which surrounds the limit cycle, is bounded. Using our earlier calculation, we

immediately obtain

V = - 8 ( x (l 0 + 3x 2

6 ) (x , 4 + 2 x 22 - 10)2

Thus V is strictly negative, except if

x,4 + 2 x 22 = I 0 or x , l 0 + 3x 2

6 = 0

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (86)

72 Fundamentals ofLyapunov Theory Chap. 3

in which case V = 0. The first equation is simply that defining the limit cycle, while the second

equation is verified only at the origin. Since both the limit cycle and the origin are invariant sets,

the set M simply consists of their union. Thus, all system trajectories starting in Q; converge

either to the limit cycle, or to the origin (Figure 3.15).

I

Figure 3.15 : Convergence to a limit cycle

Moreover, the equilibrium point at the origin can actually be shown to be unstable. However,

this result cannot be obtained from linearization, since the linearized system (x'j = x2 , x2 = 0) is

only marginally stable. Instead, and more astutely, consider the region QJQO , and note that while

the origin 0 does not belong to Q 1 0 0 , every other point in the region enclosed by the limit cycle is

in Q|QQ (in other words, the origin corresponds to a local maximum of V). Thus, while the

expression of V is the same as before, now the set M is just the limit cycle. Therefore,

reappUcation of the invariant set theorem shows that any state trajectory starting from the region

within the limit cycle, excluding the origin, actually converges to the limit cycle. In particular,

this implies that the equilibrium point at the origin is unstable. [D

Example 3.11 actually represents a very common application of the invariant settheorem: conclude asymptotic stability of an equilibrium point for systems with

negative semi-definite V. The following corollary of the invariant set theorem is morespecifically tailored to such applications:

Corollary Consider the autonomous system (3.2), with f continuous, and let V(x) bea scalar function with continuous partial derivatives. Assume that in a certainneighborhood Q of the origin

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (87)

Sect. 3.4 Lyapunov's Direct Method 73

• V(x) is locally positive definite

• V is negative semi-definite

• the set R defined by V(x) = 0 contains no trajectories of (3.2) other thanthe trivial trajectory x = 0

Then, the equilibrium point 0 is asymptotically stable. Furthermore, the largestconnected region of the form £2 (defined by V(x) < I) within £2 is a domain ofattraction of the equilibrium point.

Indeed, the largest invariant set M in R then contains only the equilibrium point 0.Note that

• The above corollary replaces the negative definiteness condition on V inLyapunov's local asymptotic stability theorem by a negative,sem/-definiteness condition on V, combined with a third condition on thetrajectories within R.

• The largest connected region of the form Q ; within £2 is a domain ofattraction of the equilibrium point, but not necessarily the whole domain ofattraction, because the function V is not unique.

• The set Q itself is not necessarily a domain of attraction. Actually, theabove theorem does not guarantee that £2 is invariant: some trajectoriesstarting in £2 but outside of the largest £2/ may actually end up outside £2.

GLOBAL INVARIANT SET THEOREMS

The above invariant set theorem and its corollary can be simply extended to a globalresult, by requiring the radial unboundedness of the scalar function V rather than theexistence of a bounded £2;.

Theorem 3.5 (Global Invariant Set Theorem) Consider the autonomous system(3.2), with f continuous, and let V(x) be a scalar function with continuous first partialderivatives. Assume that

• V(x) -> °° as !|x|| -> oo

• V(x) < 0 over the whole state space

Let R be the set of all points where V(x) = 0, and M be the largest invariant set in R.

Then all solutions globally asymptotically converge toMas(->°°.

For instance, the above theorem shows that the limit cycle convergence in

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (88)

74 Fundamentals ofLyapunov Theory Chap. 3

Example 3.13 is actually global: all system trajectories converge to the limit cycle

(unless they start exactly at the origin, which is an unstable equilibrium point).

Because of the importance of this theorem, let us present an additional (and

very useful) example.

Example 3.14: A class of second-order nonlinear systems

Consider a second-order system of the form

where b and c are continuous functions verifying the sign conditions

xb(x) > 0 for i * 0

xc(x) > 0 for x&O

The dynamics of a mass-damper-spring system with nonlinear damper and spring can be

described by equations of this form, with the above sign conditions simply indicating that the

otherwise arbitrary functions b and c actually represent "damping" and "spring" effects. A

nonlinear R-L-C (resistor-inductor-capacitor) electrical circuit can also be represented by the

above dynamic equation (Figure 3.16). Note that if the functions b and c are actually linear

( b(x) = ot[ x , c(x) = <x0 x ), the above sign conditions are simply the necessary and sufficient

conditions for the system's stability (since they are equivalent to the conditions otj > 0, ao > 0).

V

C

+

= c(x) V — X

^ — -

V =R b(x)

Figure 3.16 : A nonlinear R-L-C circuit

Together with the continuity assumptions, the sign conditions on the functions b and c imply

that b(0) = 0 and c(0) = 0 (Figure 3.17). A positive definite function for this system is

which can be thought of as the sum of the kinetic and potential energy of the system. 1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (89)

Sect. 3.4 Lyapunov's Direct Method 75

Differentiating V, we obtain

V = xx + c(x)x = -xb(x) -xc(x) + c(x)x = -xb{x) < 0

which can be thought of as representing the power dissipated in the system. Furthermore, by

hypothesis, x b(x) = 0 only if x = 0. Now x = 0 implies that

x = - c(x)

which is nonzero as long as x ^ 0. Thus the system cannot get "stuck" at an equilibrium value

other than x = 0 ; in other words, with R being the set defined by x = 0, the largest invariant set

M in R contains only one point, namely [x = 0, x = 0 ] . Use of the local invariant set theorem

indicates that the origin is a locally asymptotically stable point.

b(x) c(x)

Figure 3.17 : The functions b(x) and c(x)

Furthermore, if the integral lxc(r)dr is unbounded as |x| —> °°, then V is a radially

unbounded function and the equilibrium point at the origin is globally asymptotically stable,

according to the global invariant set theorem.

For instance, the system

3 +x5 = x4 sin2x

is globally asymptotically convergent to x = 0 (while, again, its linear approximation would be

inconclusive, even about its local stability). D

The relaxation of the positive definiteness requirement on the function V, as

compared with Lyapunov's direct method, also allows one to use a single Lyapunov-

like function to describe systems with multiple equilibria.

Example 3.15: Multimodal Lyapunov Function

Consider the system

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (90)

76 Fundamentals of Lyapunov Theory Chap. 3

! . KX+x= sin—

For this system, we study, similarly to Example 3.14, the Lyapunov function

This function has two minima, atx = ± 1; x = 0, and a local maximum in x (a saddle point in the

state-space) at x = 0; x - 0. As in Example 3.14, the time-derivative of V is (without calculations)

V = - | x 2 - l | x 4

i.e, the virtual power "dissipated" by the system. Now

V = 0 => i = 0 or i = i l

Let us consider each of these cases:

x = 0 => 3c = sin — - x * 0 except if x = 0 or x = ± 1

x = + 1 => x = 0

Thus, the invariant set theorem indicates that the system converges globally to (x = 1; x = 0) or

(x = - 1; x = 0), or to (x = 0; x = 0). The first two of these equilibrium points are stable, since they

correspond to local mimina of V (note again that linearization is inconclusive about their

stability). By contrast, the equilibrium point (x = 0; x = 0) is unstable, as can be shown from

linearization (x = (it/2 - 1) x), or simply by noticing that because that point is a local maximum of

V along the x axis, any small deviation in the x direction will drive the trajectory away from it. ID

As noticed earlier, several Lyapunov functions may exist for a given system,and therefore several associated invariant sets may be derived. The system thenconverges to the (necessarily non-empty) intersection of the invariant sets M(-, whichmay give a more precise result than that obtained from any of the Lyapunov functionstaken separately. Equivalently, one can notice that the sum of two Lyapunovfunctions for a given system is also a Lyapunov function, whose set R is theintersection of the individual sets R,-.

3.5 System Analysis Based on Lyapunov's Direct Method

With so many theorems and so many examples presented in the last section, one mayfeel confident enough to attack practical nonlinear control problems. However, thetheorems all make a basic assumption: an explicit Lyapunov function is somehow

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (91)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 77

known. The question is therefore how to find a Lyapunov function for a specificproblem. Yet, there is no general way of finding Lyapunov functions for nonlinearsystems. This is a fundamental drawback of the direct method. Therefore, faced withspecific systems, one has to use experience, intuition, and physical insights to searchfor an appropriate Lyapunov function. In this section, we discuss a number oftechniques which can facilitate the otherwise blind search of Lyapunov functions.

We first show that, not surprisingly, Lyapunov functions can be systematicallyfound to describe stable linear systems. Next, we discuss two of many mathematicalmethods that may be used to help finding a Lyapunov function for a given nonlinearsystem. We then consider the use of physical insights, which, when applicable,represents by far the most powerful and elegant way of approaching the problem, andis closest in spirit to the original intuition underlying the direct method. Finally, wediscuss the use of Lyapunov functions in transient performance analysis.

3.5.1 Lyapunov Analysis of Linear Time-Invariant Systems

Stability analysis for linear time-invariant systems is well known. It is interesting,however, to develop Lyapunov functions for such systems. First, this allows us todescribe both linear and nonlinear systems using a common language, allowing sharedinsights between the two classes. Second, as we shall detail later on, Lyapunovfunctions are "additive", like energy. In other words, Lyapunov functions forcombinations of subsystems may be derived by adding the Lyapunov functions of thesubsystems. Since nonlinear control systems may include linear components (whetherin plant or in controller), we should be able to describe linear systems in the Lyapunovformalism.

We first review some basic results on matrix algebra, since the development ofLyapunov functions for linear systems will make extensive use of quadratic forms.

SYMMETRIC, SKEW-SYMMETRIC, AND POSITIVE DEFINITEMATRICES

Definition 3.10 A square matrix M is symmetric if M = MT (in other words, ifV/,y MJ: = Mji). A square matrix M is skew-symmetric if M = — M^ (i.e. ifVi.y M^-Mji).

An interesting fact is that any square nxn matrix M can be represented as thesum of a symmetric matrix and a skew-symmetric matrix. This can be shown by thefollowing decomposition

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (92)

78 Fundamentals of Lyapunov Theory Chap. 3

.„ M + MT M - M rivl = +

2 2

where the first term on the left side is symmetric and the second term is skew-symmetric.

Another interesting fact is that the quadratic function associated with a skew-symmetric matrix is always zero. Specifically, let M be a nxn skew-symmetric matrixand x an arbitrary «xl vector. Then the definition of a skew-symmetric matrix impliesthat

Since x r M r x is a scalar, the right-hand side of the above equation can be replaced byits transpose. Therefore,

This shows that

V x , x rMx = 0 (3.16)

In designing some tracking control systems for robots, for instance, this fact is veryuseful because it can simplify the control law, as we shall see in chapter 9.

Actually, property (3.16) is a necessary and sufficient condition for a matrix Mto be skew-symmetric. This can be easily seen by applying (3.16) to the basis vectors

[ V i, e(rMse,- = 0 ] => [ V i, MH = 0 ]

and

[ V (ij), (e,- + HjfM^i + ey) = 0 ] => [V ((,;), Mu + Mtj + Mfi + MJJ = 0 ]

which, using the first result, implies that

In our later analysis of linear systems, we will often use quadratic functions ofthe form x^Mx as Lyapunov function candidates. In view of the above, each quadraticfunction of this form, whether M is symmetric or not, is always equal to a quadraticfunction with a symmetric matrix. Thus, in considering quadratic functions of the formx^Mx as Lyapunov function candidates, one can always assume, without loss ofgenerality, that M is symmetric.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (93)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 79

We are now in a position to introduce the important concept of positive definitematrices.

Definition 3.11 A square nxn matrix M is positive definite (p.d.) if

x*0 => xrMx>0

In other words, a matrix M is positive definite if the quadratic function x^Mx is apositive definite function. This definition implies that to every positive definite matrixis associated a positive definite function. Obviously, the converse is not true.

Geometrically, the definition of positive-definiteness can be interpreted assimply saying that the angle between a vector x and its image Mx is always less than90° (Figure 3.18).

Figure 3.18 : Geometric interpretation ofthe positive-definiteness of a matrix M

A necessary condition for a square matrix M to be p.d. is that its diagonalelements be strictly positive, as can be seen by applying the above definition to thebasis vectors. A famous matrix algebra result called Sylvester's theorem shows that,assuming that M is symmetric, a necessary and sufficient condition for M to be p.d. isthat its principal minors (i.e., Mjj , M^M22- M2\ Mi2> ••• > det M ) all be strictlypositive; or, equivalently, that all its eigenvalues be strictly positive. In particular, asymmetric p.d. matrix is always invertible, because the above implies that itsdeterminant is non-zero.

A positive definite matrix M can always be decomposed as

M = UrAU (3.17)

where U is a matrix of eigenvectors and satisfies UTU = I, and A is a diagonal matrixcontaining the eigenvalues of the matrix M. Let Xmin(M) denote the smallesteigenvalue of M and A,max(M) the largest. Then, it follows from (3.17) that

||2 < xTMx < Xmax(M)\\xf

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (94)

80 Fundamentals ofLyapunov Theory Chap. 3

This is due to the following three facts:

• x rMx = x rU rAUx = z rAz, where Ux = z

• Xmin(M)l< A < U M ) I

• zTz = ||x||2

The concepts of positive semi-definite, negative definite, and negative semi-definite can be defined similarly. For instance, a square n x n matrix M is said to bepositive semi-definite (p.s.d.) if

V x , xT M x > 0

By continuity, necessary and sufficient conditions for positive semi-definiteness areobtained by substituting "positive or zero" to "strictly positive" in the above conditions Ifor positive definiteness. Similarly, a p.s.d. matrix is invertible only if it is actually Ip.d. Examples of p.s.d. matrices are n x n matrices of the form M = NrN where N is a :

m x n matrix. Indeed,

V x , xT NrN x = (Nx)r (Nx) > 0

A matrix inequality of the form

M , > M 2

(where Mj and M2 are square matrices of the same dimension) means that

M, - M2 > 0

i.e., that the matrix Mj - M2 is positive definite. Similar notations apply to theconcepts of positive semi-definiteness, negative definiteness, and negative semi-definiteness.

A time-varying matrix M(f) is uniformly positive definite if

3 a > 0, V t > 0, M(0 > a I

A similar definition applies for uniform negative-definiteness of a time-varyingmatrix.

LYAPUNOV FUNCTIONS FOR LINEAR TIME-INVARIANT SYSTEMS

Given a linear system of the form x = A x , let us consider a quadratic Lyapunovfunction candidate

V = \T P x

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (95)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 81

where P is a given symmetric positive definite matrix. Differentiating the positivedefinite function V along the system trajectory yields another quadratic form

V =

where

ArP+PA=-Q

(3.18)

(3.19)

The question, thus, is to determine whether the symmetric matrix Q defined by the so-called Lyapunov equation (3.19) above, is itself p.d. If this is the case, then V satisfiesthe conditions of the basic theorem of section 3.4, and the origin is globallyasymptotically stable. However, this "natural" approach may lead to inconclusiveresult, i.e., Q may be not positive definite even for stable systems.

Example 3.17: Consider a second-order linear system whose A matrix is

A =

If we take P = I, then

4

- 1 2

0 - 4

- 4 -24

The matrix Q is not positive definite. Therefore, no conclusion can be drawn from the Lyapunov

function on whether the system is stable or not. C]

A more useful way of studying a given linear system using scalar quadraticfunctions is, instead, to derive a positive definite matrix P from a given positivedefinite matrix Q, i.e.,

• choose a positive definite matrix Q

• solve for P from the Lyapunov equation (3.19)

• check whether P is p.d

If P is p.d., then x^Px is a Lyapunov function for the linear system and globalasymptotical stability is guaranteed. Unlike the previous approach of going from agiven P to a matrix Q, this technique of going from a given Q to a matrix P alwaysleads to conclusive results for stable linear systems, as seen from the followingtheorem.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (96)

82 Fundamentals of Lyapunov Theory Chap. 3

Theorem 3.6 A necessary and sufficient condition for a LTl system x = A x to bestrictly stable is that, for any symmetric p.d. matrix Q, the unique matrix P solution ofthe Lyapunov equation (3.19) be symmetric positive definite.

Proof: The above discussion shows that the condition is sufficient, thus we only need to show

that it is also necessary. We first show that given any symmetric p.d. matrix Q, there exists a

symmetric p.d. matrix P verifying (3.19). We then show that for a given Q, the matrix P is

actually unique.

Let Q be a given symmetric positive definite matrix, and let

P = f exp(Ar t) Q exp(A r) dt (3.20)

One can easily show that this integral exists if and only if A is strictly stable. Also note that the

matrix P thus defined is symmetric and positive definite, since Q is. Furthermore, we have

oo= J [ A r exp(Ar 0 Q exp(A t) + exp(Ar t) Q exp(A t)A]dt

= A r P + P A

where the first equality comes from the stability of A (which implies that exp (A°°) = 0), the

second from differentiating the exponentials explicitly, and the third from the fact that A is

constant and therefore can be taken out of the integrals.

The uniqueness of P can be verified similarly by noting that another solution Pj of the

Lyapunov equation would necessarily verify

ooPi = - ] 4exp(Ar0P1exp(A()]

oo= - J exp(AT f) ( A r P, + P, A ) exp(A I) ] dt

oo= I exp(Ar t) Q exp(A t)dt = V

J0

An alternate proof of uniqueness is the elegant original proof given by Lyapunov, which

makes direct use of fundamental algebra results. The Lyapunov equation (3.19) can be

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (97)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 83

interpreted as defining a linear map from the n2 components of the matrix P to the n2 components

of the matrix Q, where P and Q are arbitrary (not necessarily symmetric p.d.) square matrices.

Since (3.20) actually shows the existence of a solution P for any square matrix Q, the range of

this linear map is full, and therefore its null-space is reduced to 0. Thus, for any Q, the solution P

is unique. D

The above theorem shows that any positive definite matrix Q can be used todetermine the stability of a linear system. A simple choice of Q is the identity matrix.

Example 3.18: Consider again the second-order system of Example 3.17. Let us take Q = I anddenote P by

P =Pll Pl2

Pl\ P22

where, due to the symmetry of P, p2 i = Pu- Then the Lyapunov equation is

4

- 8

- 1 2

P l l

Pl2

P\2

P22

- 1

- 1

whose solution is

p n = 5 / 1 6 , pl2=P22=

The corresponding matrix

16

5 1

1 1

is positive definite, and therefore the linear system is globally asymptotically stable. Note that we

have solved for P directly, without using the more cumbersome expression (3.20). CD

Even though the choice Q = I is motivated by computational simplicity, it has asurprising property: the resulting Lyapunov analysis allows us to get the best estimateof the state convergence rate, as we shall see in section 3.5.5.

3.5.2 Krasovskii's Method

Let us now come back to the problem of finding Lyapunov functions for generalnonlinear systems. Krasovskii's method suggests a simple form of Lyapunov function

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (98)

84 Fundamentals ofLyapunov Theory Chap. 3

candidate for autonomous nonlinear systems of the form (3.2), namely, V = f f. Thebasic idea of the method is simply to check whether this particular choice indeed leadsto a Lyapunov function.

Theorem 3.7 (Krasovskii) Consider the autonomous system defined by (3.2), withthe equilibrium point of interest being the origin. Let A(x) denote, the Jacobian matrixof the system, i.e.,

A(x) = —d x

// the matrix F = A + \T is negative definite in a neighborhood Q., then theequilibrium point at the origin is asymptotically stable. A Lyapunov function for thissystem is

V(x) = f{x) f(x)

// Q. is the entire state space and, in addition, V(x) —> °° as \\x\\ -> °o, then theequilibrium point is globally asymptotically stable.

Proof: First, let us prove that the negative definiteness of F implies that f(x) ^ 0 for x * 0.

Since the square matrix F(x) is negative definite for non-zero x, one can show that the

Jacobian matrix A(x) is invertible, by contradiction. Indeed, assume that A is singular. Then one

can find a non-zero vector y0 such that A(x)y0 = 0. Since

the singularity of A implies that yo^A yg = 0, which contradicts the assumed negative definiteness

ofF.

The invertibility and continuity of A guarantee that the function f(x) can be uniquely

inverted. This implies that the dynamic system (3.2) has only one equilibrium point in £i

(otherwise different equilibrium points would correspond to the same value of f), i.e., that f(x) ^ 0

for x * 0.

We can now show the asymptotic stability of the origin. Given the above result, the scalar

function V(x) = f^(x) f(x) is positive definite. Using the fact that f = A f, the derivative of V can be

written

V(x) = frf + frf = f r A f + f r A r f = f rFf

The negative definiteness of F implies the negative definiteness of V. Therefore, according to

Lyapunov's direct method, the equilibrium state at the origin is asymptotically stable. The global

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (99)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 85

asymptotic stability of the system is guaranteed by the global version of Lyapunov's direct

method. D

Let us illustrate the use of Krasovskii's theorem on a simple example.

Example 3.19: Consider the nonlinear system

^2 = 2jtj — 6^2 — 2x 2

We have

-6

2 - 6 - 6 * 2•2

F = A + A r =- 1 2 4

4 - 1 2 - 1 2 *

The matrix F is easily shown to be negative definite over the whole state space. Therefore, the

origin is asymptotically stable, and a Lyapunov function candidate is

V(x) = fr(x)f(x) = ( - 6 * , + 2x2)2 + ( 2 x , - 6 x 2 - 2 x 23 ) 2

Since V(x) -> °° as ||x|| —> °° , the equilibrium state at the origin is globally asymptotically

stable. •

While the use of the above theorem is very straightforward, its applicability islimited in practice, because the Jacobians of many systems do not satisfy the negativedefiniteness requirement. In addition, for systems of high order, it is difficult to checkthe negative definiteness of the matrix F for all x.

An immediate generalization of Krasovskii's theorem is as follows:

Theorem 3.8 (Generalized Krasovskii Theorem) Consider the autonomous systemdefined by (3.2), with the equilibrium point of interest being the origin, and let A(x)denote the Jacobian matrix of the system. Then, a sufficient condition for the origin tobe asymptotically stable is that there exist two symmetric positive definite matrices Pand Q, such that Vx ^ 0, the matrix

F(x) = A 7 P + P A + Q

is negative semi-definite in some neighborhood Q of the origin. The functionV(x) = f^Pf is then a Lyapunov function for the system. If the region Q is the wholestate space, and if in addition, V(x) -^ °° as \\x\\ —> °°, then the system is globallyasymptotically stable.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (100)

86 Fundamentals of Lyapunov Theory Chap. 3

Proof: This theorem can be proven similarly. The positive definiteness of V(x) can be derived as

before. Furthermore, the derivative of V can be computed as

V = — f(x) = fTPA(x)f + frPAr(x)Pf = frFf-frQf3x

Because F is negative semi-definite and Q is positive definite, V is negative definite and the

equilibrium point at the origin is asymptotically stable. If V(x) —> °° as ||x|| —> °° , the global

version of Lyapunov's direct method indicates the global asymptotic stability of the system. D

3.5.3 The Variable Gradient Method

The variable gradient method is a formal approach to constructing Lyapunovfunctions. It involves assuming a certain form for the gradient of an unknownLyapunov function, and then finding the Lyapunov function itself by integrating theassumed gradient. For low order systems, this approach sometimes leads to thesuccessful discovery of a Lyapunov function.

To start with, let us note that a scalar function V(x) is related to its gradient Wby the integral relation

V(x) = Cwdx

where W = {dV/dx^, ,3V/dxn}T. In order to recover a unique scalar function V

from the gradient VV, the gradient function has to satisfy the so-called curl conditions

a v V: a v V:^r^ = - ~ (/,;= i,2,...,n)

a Xj 3 Xj

Note that the ith component VV,- is simply the directional derivative d V/3 xi . Forinstance, in the case n = 2, the above simply means that

dVVy d V V2

The principle of the variable gradient method is to assume a specific form forthe gradient VV, instead of assuming a specific form for the Lyapunov function Vitself. A simple way is to assume that the gradient function is of the form

VV; = j ^ a l j X j (3.21)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (101)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 87

where the a,-.'s are coefficients to be determined. This leads to the followingprocedure for seeking a Lyapunov function V :

• assume that VV is given by (3.21) (or another form)

• solve for the coefficients at: so as to satisfy the curl equations

• restrict the coefficients in (3.21) so that V is negative semi-definite (atleast locally)

• compute V from VV by integration;

• check whether V is positive definite

Since satisfaction of the curl conditions implies that the above integration result isindependent of the integration path, it is usually convenient to obtain V by integratingalong a path which is parallel to each axis in turn, i.e.,

V(x) = f ' W ^ . O . - . - . O ) dx{ + f 2 VV2(xl,x2,0,...,0)dx2 + ...o o

Example 3.20: Let us use the variable gradient method to find a Lyapunov function for the

nonlinear system

i>> = — 2*2

We assume that the gradient of the undetermined Lyapunov function has the following form

The curl equation

dx2

al2 + .

is

dal2

*2 3x 2 ° 2 1

3a 2 1

X] 3 * ,

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (102)

88 Fundamentals of Lyapunov Theory Chap. 3

If the coefficients are chosen to be

an=a22=\,an = a2l=0

which leads to

W , = x , W 2 = x2

then V can be computed as

V = VV k = - 2x,2 - 2x22( 1 -x y x 2 )

Thus, V is locally negative definite in the region (1 - xxx2) > 0 . The function V can be computed

as

-x. -x, xt2 + x-?-

V(x)= f x,dx, + f x2dx2 = (3.22)

This is indeed positive definite, and therefore the asymptotic stability is guaranteed.

Note that (3.22) is not the only Lyapunov function obtainable by the variable gradient

method. For example, by taking

a,, = l, an = x22

«21=3*22 • a22=3

we obtain the positive definite function

V = ^ i - + L 22 + x,x2

3 (3.23)

whose derivative is

V = - 2x,2 - 6x22 - 2x2

2 (x,x2 - 3x,2x22)

One easily verifies that V is a locally negative definite function (noting that the quadratic terms

are dominant near the origin), and therefore, (3.23) represents another Lyapunov function for the

system. uJ

3.5.4 Physically Motivated Lyapunov Functions

The Lyapunov functions in the above sections 3.5.1-3.5.3, and in a number of

examples earlier in section 3.4, have been obtained from a mathematical point of

view, i.e., we examined the mathematical features of the given differential equations

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (103)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 89

and searched for Lyapunov function candidates V that can make V negative. We did

not pay much attention to where the dynamic equations came from and what

properties the physical systems had. However, this purely mathematical approach,

though effective for simple systems, is often of little use for complicated dynamic

equations. On the other hand, if engineering insight and physical properties are

properly exploited, an elegant and powerful Lyapunov analysis may be possible for

very complex systems.

Example 3.21: Global asymptotic stability of a robot position controller

A fundamental task in robotic applications is for robot manipulators to transfer objects from one

point to another, the so-called robot position control problem. In the last decade, engineers had

been routinely using P.D. (proportional plus derivative) controllers to control robot arms.

However, there was no theoretical justification for the stability of such control systems, because

the dynamics of a robot is highly nonlinear.

A robot arm consists a number of links connected by rotational or translational joints, with

the last link equipped with some end-effector (Figure 3.19). The dynamics of an n-link robot arm

can be expressed by a set of n equations,

H(q) q + b(q, q) + g(q) = T (3.24)

where q is an ^-dimensional vector describing the joint positions of the robot, x is the vector of

input torques, g is the vector of gravitational torques, b represents the Coriolis and centripetal

forces caused by the motion of the links, and H the nxn inertia matrix of the robot arm. Consider

a controller simply composed of a P.D. term and a gravity compensation term

(3.25)

where KD and Kp are constant positive definite nxn matrices. It is almost impossible to use

trial-and error to search for a Lyapunov function for the closed loop dynamics defined by (3.24)

and (3.25), because (3.24) contains hundreds of terms for the 5-Hnk or 6-link robot arms

commonly found in industry. Therefore, it seems very difficult to show that q —> 0 and q —» 0.

With the aid of physical insights, however, a Lyapunov function can be successfully found

for such complex robotic systems. First, note that the inertia matrix H(q) is positive definite for

any q. Second, the P.D. control term can be interpreted as mimicking a combination of dampers

and springs. This suggests the following Lyapunov function candidate

V = I [q rHq+ qrKpq]

where the first term represents the kinetic energy of the manipulator, and the second term denotes

the "artificial potential energy" associated with the virtual spring in the control law (3.25).

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (104)

90 Fundamentals of Lyapunov Theory

joint 2

Chap. 3

base(link 0)

Figure 3.19 : A robot manipulator

In computing the derivative of this function, one can use the energy theorem in mechanics,

which states that the rate of change of kinetic energy in a mechanical system is equal to the power

provided by the external forces. Therefore,

Substitution of the control law (3.25) in the above equation then leads to

Since the arm cannot get "stuck" at any position such that q # 0 (which can be easily shown by

noting that acceleration is non-zero in such situations), the robot arm must settle down at q = 0

and q = 0, according to the invariant set theorem. Thus, the system is actually globally

asymptotically stable. O

Two lessons can be learned from this practical example. The first is that oneshould use as many physical properties as possible in analyzing the behavior of asystem. The second lesson is that physical concepts like energy may lead us to someuniquely powerful choices of Lyapunov functions. Physics will play a major role inthe development of the multi-input nonlinear controllers of chapter 9.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (105)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 91

3.5.5 Performance Analysis

In the preceding sections, we have been primarily concerned with using Lyapunovfunctions for stability analysis. But sometimes Lyapunov functions can further provideestimates of the transient performance of stable systems. In particular, they can allowus to estimate the convergence rate of asymptotically stable linear or nonlinearsystems. In this section, we first derive a simple lemma on differential inequalities.We then show how Lyapunov analysis may be used to determine the convergencerates of linear and nonlinear systems.

A SIMPLE CONVERGENCE LEMMA

Lemma: If a real function W(t) satisfies the inequality

W(t) + aW(t) < 0 (3.26)

where a is a real number. Then

W(t)<W(0)e~at

Proof: Let us define a function Z(t) by

Z(t) = W + aW (3.27)

Equation (3.26) implies that Z(t) is non-positive. The solution of the first-order equation (3.27) is

W(t) = W(0)e~a'+i e~a(;"r)Z{t)dr* o

Because the second term in the right-hand-side of the above equation is non-positive, one has

W(t)<W(0)e-at •

The above lemma implies that, if W is a non-negative function, the satisfactionof (3.26) guarantees the exponential convergence of W to zero. In using Lyapunov'sdirect method for stability analysis, it is sometimes possible to manipulate V into theform of (3.26). In such a case, the exponential convergence of V and the convergencerate can be inferred and, in turn, the exponential convergence rate of the state maythen be determined. In later chapters, we will provide examples of using this lemmafor estimating convergence rates of nonlinear or adaptive control systems.

ESTIMATING CONVERGENCE RATES FOR LINEAR SYSTEMS

Now let us evaluate the convergence rate of a stable linear system based on theLyapunov analysis described in section 3.5.1. Let us denote the largest eigenvalue of

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (106)

92 Fundamentals ofLyapunov Theory Chap. 3

the matrix P by A,max(P), the smallest eigenvalue of Q by A.min(Q), and their ratio^•mitfQir^mwfP) by Y- The positive definiteness of P and Q implies that these scalarsare all strictly positive. Since matrix theory shows that

we have

x r Q x ;>

This and (3.18) imply that

V<-yV

This, according to lemma, means that

This, together with the fact x^Px > X,mln(P)||x(f)ll2 , implies that the state x convergesto the origin with a rate of at least y/2.

One might naturally wonder how this convergence rate estimate varies with thechoice of Q, and how it relates to the familiar notion of dominant pole in linear theory.An interesting result is that the convergence rate estimate is largest for Q = I. Indeed,let P o be the solution of the Lyapunov equation corresponding to Q = I :

and let P be the solution corresponding to some other choice of Q

Without loss of generality, we can assume that ^m,>,(Qi) = 1, since rescaling Qj willrescale P by the same factor, and therefore will not affect the value of thecorresponding y. Subtracting the above two equations yields

Now since X.mi-n(Qj) = 1 = X,max(I), the matrix (Q[ - 1 ) is positive semi-definite, andhence the above equation implies that (P - Po) is positive semi-definite. Therefore

Since A,m;n(Qj) = 1 = A,m;n(I), the convergence rate estimate

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (107)

Sect. 3.5 System Analysis Based on Lyapunov's Direct Method 93

corresponding to Q = I is larger than (or equal to) that corresponding to Q = Qj.

If the stable matrix A is symmetric, then the meaning of this "optimal" value ofy, corresponding to the choice Q = I, can be interpreted easily. Indeed, alleigenvalues of A are then real, and furthermore A is diagonalizable, i.e., there exists achange of state coordinates such that in these coordinates A is diagonal. Oneimmediately verifies that, in these coordinates, the matrix P = - 1/2 A~ ' verifies theLyapunov equation for Q = I, and that therefore the corresponding y/2 is simply theabsolute value of the dominant pole of the linear system. Furthermore, y is obviouslyindependent of the choice of state coordinates.

ESTIMATING CONVERGENCE RATES FOR NONLINEAR SYSTEMS

The estimation of convergence rate for nonlinear systems also involves manipulatingthe expression of V so as to obtain an explicit estimate of V. The difference lies in that,for nonlinear systems, V and V are not necessarily quadratic functions of the states.

Example 3.22: Consider again the system in Example 3.8. Given the chosen Lyapunov function

candidate V= ||x||2, the derivative V, can be written

V=2V(V- 1)

that is,

dV- = -2dt

V(l -V)

The solution of this equation is easily found to be

9 /CLP~

V(x)= ae

1 +ae~2'

where

1 - V(0)

If ||x(0)||2 = V(0) < I, i.e., if the trajectory starts inside the unit circle, then a > 0, and

V(t)<ae~2'

This implies that the norm ||x(/)|| of the state vector converges to zero exponentially, with a rate

of 1.

However, if the trajectory starts outside the unit circle, i.e., if V(0) > I, then a < 0 , so that

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (108)

94 Fundamentals of Lyapunov Theory Chap. 3

V(i) and therefore ||x|| tend to infinity in a finite time (the system is said to exhibit finite escape

time, or "explosion"). I—I

3.6 Control Design Based on Lyapunov's Direct Method

The previous sections have been concerned with using Lyapunov's direct method forsystem analysis. In doing the analysis, we have implicitly presumed that certaincontrol laws have been chosen for the systems. However, in many control problems,the task is to find an appropriate control law for a given plant. In the following, webriefly discuss how to apply Lyapunov's direct method for designing stable controlsystems. Many of the controller design methods we will describe in chapters 6-9 areactually based on Lyapunov concepts.

There are basically two ways of using Lyapunov's direct method for controldesign, and both have a trial and error flavor. The first technique involveshypothesizing one form of control law and then finding a Lyapunov function to justifythe choice. The second technique, conversely, requires hypothesizing a Lyapunovfunction candidate and then finding a control law to make this candidate a realLyapunov function.

We saw an application of the first technique in the robotic P.D. controlexample in section 3.5.4, where a P.D. controller is chosen based on physical intuition,and a Lyapunov function is found to show the global asymptotic convergence of theresulting closed-loop system. The second technique can be illustrated on thefollowing simple example.

Example 3.23: Regulator Design

Consider the problem of stabilizing the system

i.e., to bring it to equilibrium at x = 0. Based on Example 3.14, it is sufficient to choose a

continuous control law u of the form

where

i ( i 3 + u1(i))<0 for i

x(x2 - u2(x)) > 0 for

The above inequalities also imply that globally stabilizing controllers can be designed even in

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (109)

Sect. 3.7 Summary 95

the presence of some uncertainty on the dynamics. For instance, the system

where the constants o^ and a2 are unknown, but such that CC[ > - 2 and lo^l < 5, can be globally

stabilized using the control law

u = -2p-5x\x\ •

For some classes of nonlinear systems, systematic design procedures have beendeveloped based on these above two techniques, as can be seen in chapter 7 on thesliding control design methodology, in chapter 8 on adaptive control, and in chapter 9on physically-motivated designs.

Finally, it is important to notice that, just as a nonlinear system may be globallyasymptotically stable while its linear approximation is only marginally stable, anonlinear system may be controllable while its linear approximation is not. Considerfor instance the system

x + x5 - x2 u

This system can be made to converge asymptotically to zero simply by letting u = - x.However, its linear approximation at x = 0 and u = 0 is x ~ 0 , and therefore isuncontrollable!

3.7 Summary

Stability is a fundamental issue in control system analysis and design. Variousconcepts of stability, such as Lyapunov stability, asymptotic stability, exponentialstability, and global asymptotic or exponential stability, must be defined in order toaccurately characterize the complex and rich stability behaviors exhibited by nonlinearsystems.

Since analytical solutions of nonlinear differential equations usually cannot beobtained, the two methods of Lyapunov are of major importance in the determinationof nonlinear system stability.

The linearization method is concerned with the small motion of nonlinearsystems around equilibrium points. It is mainly of theoretical value, justifying the useof linear control for the design and analysis of weakly nonlinear physical systems.

The direct method, based on so-called Lyapunov functions, is not restricted tosmall motions. In principle, it is applicable to essentially all dynamic systems, whether

L

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (110)

96 Fundamentals of Lyapunov Theory Chap. 3

linear or nonlinear, continuous-time or discrete-time , of finite or infinite order, and insmall or large motion. However, the method suffers from the common difficulty offinding a Lyapunov function for a given system. Since there is no generally effectiveapproach for finding Lyapunov functions, one has to use trial-and-error, experience,and intuition to search for appropriate Lyapunov functions. Physical properties (suchas energy conservation) and physical insights may be exploited in formulatingLyapunov functions, and may lead to uniquely powerful choices. Simple mathematicaltechniques, such as, e.g., Krasovskii's method or the variable gradient method, canalso be of help.

Generally, the application of Lyapunov theory to controller design is moreeasily rewarding. This is because, in design, one often has the freedom to deliberatelymodify the dynamics (by designing the controller) in such a way that a chosen scalarfunction becomes a Lyapunov function for the closed-loop system. In the second partof the book, we will see many applications of Lyapunov theory to the construction ofeffective nonlinear control systems.

3.8 Notes and References

In Lyapunov's original work [Lyapunov, 1892], the linearization method (which, today, is

sometimes incorrectly referred to as the first method) is simply given as an example of application of

the direct (or second) method. The first method was the so-called method of exponents, which is

used today in the analysis of chaos.

Many references can be found on Lyapunov theory, e.g., [La Salle and Lefschetz, 1961] (on

which the invariant set theorems are based), [Kalman and Bertram, 1960; Hahn, 1967; Yoshizawa,

1966]. An inspiring discussion of the role of scalar summarizing functions in science, along with a

very readable introduction to elementary Lyapunov theory, can be found in [Luenberger, 1979],

Examples 3.1 and 3.13 are adapted from [Luenberger, 1979]. Examples 3.3 and 3.8 are

adapted from [Vidyasagar, 1978]. The counterexample in Figure 3.12 is from [Hahn, 1967].

Example 3.14 is adapted from [La Salle and Lefschetz, 1961; Vidyasagar, 1978]. The variable

gradient method in subsection 3.5.3 is adapted from [Ogata, 1970]. The robotic example of section

3.5.4 is based on [Takegaki and Arimoto, 1981]. The remark on the "optimal" choice of Q in section

3.5.5 is from [Vidyasagar, 1982]. A detailed study of Krasovskii's theorems can be found in

[Krasovskii, 1963].

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (111)

Sect. 3.9 Exercises 97

3.9 Exercises

3.1 The norm used in the definitions of stability need not be the usual Euclidian norm. If the state-

space is of finite dimension n (i.e., the state vector has n components), stability and its type are

independent of the choice of norm (all norms are "equivalent"), although a particular choice of norm

may make analysis easier. For n = 2 , draw the unit balls corresponding to the following norms:

(i) | | x | | 2 = (x l)2 + (jf2)

2 (Euclidian norm)

(ii) | | x | | 2 = (x,)2 + 5(x2)2

(iii) 11x11 = 1*, |+ |*21

(iv) | | x | | = S u p ( | * , | , | * 2 | )

Recall that a ball B(xo , R), of center x0 and radius /?, is the set of x such that || x - x j | < R , and that

the unit ball is B(0, 1).

3.2 For the following systems, find the equilibrium points and determine their stability. Indicate

whether the stability is asymptotic, and whether it is global.

(a) x = — x 3 + sin *

(b) i = (5-x)5

(c) x + x 5 + x1 =JC2 s in 8xcos 23x

(d) x + (x- l ) 4 i 7 +x 5 =* 3 s in 3 x

(e) * + (*- l ) 2 x 7 + x = sin(itx/2)

3.3 For the Van der Pol oscillator of Example 3.3, demonstrate the existence of a limit cycle using

the linearization method.

3.4 This exercise, adapted from [Hahn, 1967], provides an example illustrating the motivation of

the radial unboundedness condition in Theorem 3.3. Consider the second-order system

2(*, +x2)

with z = 1 +X|2. On the hyperbola x2 = 2/(xj — " 2 ), the system trajectory slope is

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (112)

98 Fundamentals of Lyapunov Theory Chap. 3

x 2 _ i

while the slope of the hyperbola is

dxx i -23l2xl + x^/2

Note that for x, > ~\j~2 , the first expression is larger than the second, implying that the trajectories

cannot cut the branch of the hyperbola which lies in the first quadrant, in the direction toward the

axes (since on the hyperbola we have i , > 0 if JCJ > ~NJ2 ). Thus, there are trajectories which do not

tend toward the origin, indicating the lack of global asymptotic stability. Use the scalar function

x 2

V(x) = - L + x22

to analyze the stability of the above system.

3.5 Determine regions of attraction of the pendulum, using as Lyapunov functions the pendulum's

total energy, and the modified Lyapunov function of page 67. Comment on the two results.

3.6 Show that given a constant matrix M and any time-varying vector x, the time-derivative of the

scalar x r M x can be written

- x r M x = xT(M + MT )x = x r ( M + M r ) xdt

and that, if M is symmetric, it can also be written

- x T M x = 2 x r M x = 2 x r M xdt

3.7 Consider an n x n matrix M of the form M = N rN , where N i s a m X n matrix. Show that M is

p.d. if, and only if, m > n and N has full rank.

3.8 Show that if M is a symmetric matrix such that

V x , x r M x = 0

then M = 0.

3.9 Show that if symmetric p.d. matrices P and Q exist such that

A T P + P A + 2X.P = - Q

then all the eigenvalues of A have a real part strictly less than - X. (Adapted from [Luenberger,

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (113)

Sect. 3.9 Exercises 99

1979].)

3.10 Consider the system

A , y + A 2 j + A 3 y = 0

where the 2n x 1 vector x = [yT yT]T is the state, and the n x n matrices A.j ace all symmetric

positive definite. Show that the system is globally asymptotically stable, with 0 as a unique

equilibrium point.

3.11 Consider the system

x = A x y = c^x

Use the invariant set theorem to show that if the system is observable, and if there exists a symmetric

p.d. matrix P such that

ATP + l

then the system is asymptotically stable.

Can the result be derived using the direct method? (Adapted from [Luenberger, 1979].)

3.12 Use Krasovskii's theorem to justify Lyapunov's linearization method.

3.13 Consider the system

x = 4x2y-fl(x)(x2

y = -2xi-f2(y)(x2

where the continuous functions / j and / 2 have the same sign as their argument. Show that the

system tends towards a limit cycle independent of the explicit expressions of/| and/2 .

3.14 The second law of thermodynamics states that the entropy of an isolated system can only

increase with time. How does this relate to the notion of a Lyapunov function?

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (114)

Chapter 4Advanced Stability Theory

In the previous chapter, we studied Lyapunov analysis of autonomous systems. Inmany practical problems, however, we encounter non-autonomous systems. Forinstance, a rocket taking off is a non-autonomous system, because the parametersinvolved in its dynamic equations, such as air temperature and pressure, vary withtime. Furthermore, as discussed earlier, determining the stability of a nominal motionfor an autonomous system requires the stability analysis of an equivalent non-autonomous system around an equilibrium point. Therefore, stability analysistechniques for non-autonomous systems must be developed. This constitutes the majortopic of this chapter.

After extending the concepts of stability and the Lyapunov stability theorems tonon-autonomous systems, in sections 4.1 and 4.2, we discuss a number of interestingrelated topics in advanced stability theory. Some Lyapunov theorems for concludinginstability of nonlinear systems are provided in section 4.3. Section 4.4 discusses anumber of so-called converse theorems asserting the existence of Lyapunov functions.Besides their theoretical interest, these existence theorems can be valuable in somenonlinear system analysis and design problems. In section 4.5, we describe a verysimple mathematical result, known as Barbalat's lemma, which can be convenientlyused to solve some asymptotic stability problems beyond the treatment of Lyapunovstability theorems, and which we shall use extensively in chapters 8 and 9. In section4.6, we discuss the so-called positive real linear systems and their unique properties,which shall be exploited later in the book, particularly in chapter 8. Section 4.7describes passivity theory, a convenient way of interpreting, representing, and

100

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (115)

Sect. 4.1 Concepts of Stability for Non-Autonomous Systems 101

combining Lyapunov or Lyapunov-like functions. Section 4.8 discusses a special classof nonlinear systems which can be systematically treated by Lyapunov analysis.Section 4.9 studies some non-Lyapunov techniques which can be used to establishboundedness of signals in nonlinear systems. Finally, section 4.10 discussesmathematical conditions for the existence and unicity of solutions of nonlineardifferential equations.

4.1 Concepts of Stability for Non-Autonomous Systems

The concepts of stability for non-autonomous systems are quite similar to those ofautonomous systems. However, due to the dependence of non-autonomous systembehavior on initial time to, the definitions of these stability concepts include t0

explicitly. Furthermore, a new concept, uniformity, is necessary to characterize non-autonomous systems whose behavior has a certain consistency for different values ofinitial time t0. In this section, we concisely extend the stability concepts forautonomous systems to non-autonomous systems, and introduce the new concept ofuniformity.

EQUILIBRIUM POINTS AND INVARIANT SETS

For non-autonomous systems, of the form

x = f(x, t) (4.1)

equilibrium points x are defined by

f(x*,0 = 0 Vt>to (4.2)

Note that this equation must be satisfied V t> to , implying that the system should beable to stay at the point x* all the time. For instance, one easily sees that the lineartime-varying system

x = A(r)x (4.3)

has a unique equilibrium point at the origin 0 unless A(?) is always singular.

Example 4.1: The system

k = -*££- (4.4)l+x2

has an equilibrium point at x = 0. However, the system

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (116)

102 Advanced Stability Theory Chap. 4

x = - HQl. + fc(r) (4.5)1 + x 2

with b(t) ^ 0, does not have an equilibrium point. It can be regarded as a system under external

input or disturbance b(f). Since Lyapunov theory is mainly developed for the stability of nonlinear

systems with respect to initial conditions, such problems of forced motion analysis are more

appropriately treated by other methods, such as those in section 4.9. LJ

The definition of invariant set is the same for non-autonomous systems as forautonomous systems. Note that, unlike in autonomous systems, a system trajectory isgenerally not an invariant set for a non-autonomous system (Exercise 4.1).

EXTENSIONS OF THE PREVIOUS STABILITY CONCEPTS

Let us now extend the previously defined concepts of stability, instability, asymptoticstability, and exponential stability to non-autonomous systems. The key in doing so isto properly include the initial time t0 in the definitions.

Definition 4.1 The equilibrium point 0 is stable at t0 if for any R>0, there exists apositive scalar r{R,t0) such that

\\x(to)\\<r => \\x(t)\\<R Vr>fo (4.6)

Otherwise, the equilibrium point 0 is unstable.

Again, the definition means that we can keep the state in a ball of arbitrarily smallradius R by starting the state trajectory in a ball of sufficiently small radius r.Definition 4.1 differs from definition 3.3 in that the radius r of the initial ball maydepend on the initial time t0.

The concept of asymptotic stability can also be defined for non-autonomous

systems.

Definition 4.2 The equilibrium point 0 is asymptotically stable at time to if

• it is stable

• 3 r(t0) > 0 such that \\ x(t0) \\ < r{to) => || x(t) || -» 0 as t -> oo

Here, the asymptotic stability requires that there exists an attractive region for everyinitial time t0 . The size of attractive region and the speed of trajectory convergencemay depend on the initial time t0.

The definitions of exponential stability and global asymptotic stability are alsostraightforward.

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (117)

Sect. 4.1 Concepts of Stability for Non-Autonomous Systems 103

Definition 4.3 The equilibrium point 0 is exponentially stable if there exist twopositive numbers, a and X, such that for sufficiently small x(to) ,

Definition 4.4 The equilibrium point 0 is globally asymptotically stable if\/ x(/0)

x(f) -> 0 as t —> °°

Example 4.2: A first-order linear time-varying system

Consider the first-order system

x(t) = -a(t)x(t)

Its solution is

x(0 = x(ro)exp[-f a(r)dr]

Thus, the system is stable if a(t) > 0, V l > t0 . It is asymptotically stable if f°° a(r) dr = + °° . It

is exponentially stable if there exists a strictly positive number T such that V / > 0,

f'+ a(r) dr > y, with y being a positive constant.

For instance,

• The system x = -x/( 1 + t)2 is stable (but not asymptotically stable).

• The system x = - x/(l + t) is asymptotically stable.

• The system x = — tx is exponentially stable.

Another interesting example is the system

1 + sinxz

whose solution can be expressed as

' - 140

Since

r

=x(ro)exp[f -=—Ji 1 +sinxz(r)

1 , . <->o-dr >i I +sinx2(V) 2

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (118)

104 Advanced Stability Theory Chap. 4

the system is exponentially convergent with rate 1/2 . d | |

UNIFORMITY IN STABILITY CONCEPTS

The previous concepts of Lyapunov stability and asymptotic stability for non-autonomous systems both indicate the important effect of initial time. In practice, it isusually desirable for the system to have a certain uniformity in its behavior regardlessof when the operation starts. This motivates us to consider the definitions of uniformstability and uniform asymptotic stability. As we shall see in later chapters, non-autonomous systems with uniform properties have some desirable ability to withstanddisturbances. It is also useful to point out that, because the behavior of autonomoussystems is independent of the initial time, all the stability properties of an autonomoussystem are uniform.

Definition 4.5 The equilibrium point 0 is locally uniformly stable if the scalar r inDefinition 4.1 can be chosen independently ofto , i.e., ifr = r(R).

The intuitive reason for introducing the concept of uniform stability is to ruleout systems which are "less and less stable" for larger values of t0 . Similarly, thedefinition of uniform asymptotic stability also intends to restrict the effect of the initialtime t0 on the state convergence pattern.

Definition 4.6 The equilibrium point at the origin is locally uniformlyasymptotically stable if

• it is uniformly stable

• There exists a ball of attraction BR , whose radius is independent of t0 ,such that any system trajectory with initial states in BR converges to 0uniformly in t0

By uniform convergence in terms of t0, we mean that for all R\ and R2 satisfying0 < R2 < /?! < Ro , 3 T{Rb R2) > 0, such that, V tg > 0,

\\«S0)\\<R\ => Hx(0ll<K2 Vr>fo + r(/?!,/?2)

i.e., the state trajectory, starting from within a ball BR , will converge into a smallerball BR after a time period T which is independent of t0 .

By definition, uniform asymptotic stability always implies asymptotic stability.The converse is generally not true, as illustrated by the following example.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (119)

Sect. 4.2 Lyapunov Analysis of Non-Autonomous Systems 105

Example 4.3: Consider the first-order system

xx = —

1 +tThis system has the general solution

4 0 = - r ^ X{tn)I +t "

This solution asymptotically converges to zero. But the convergence is not uniform. Intuitively,

this is because a larger to requires a longer time to get close to the origin. D

Using Definition 4.3, one can easily prove that exponential stability alwaysimplies uniform asymptotic stability.

The concept of globally uniformly asymptotic stability can be defined byreplacing the ball of attraction B^ by the whole state space.

4.2 Lyapunov Analysis of Non-Autonomous Systems

We now extend the Lyapunov analysis results of chapter 3 to the stability analysis ofnon-autonomous systems. Although many of the ideas in chapter 3 can be similarlyapplied to the non-autonomous case, the conditions required in the treatment of non-autonomous systems are more complicated and more restrictive. We start with thedescription of the direct method. We then apply the direct method to the stabilityanalysis of linear time-varying systems. Finally, we discuss the linearization methodfor non-autonomous nonlinear systems.

4.2.1 Lyapunov's Direct Method for Non-Autonomous Systems

The basic idea of the direct method, i.e., concluding the stability of nonlinear systemsusing scalar Lyapunov functions, can be similarly applied to non-autonomous systems.Besides more mathematical complexity, a major difference in non-autonomoussystems is that the powerful La Salle's theorems do not apply. This drawback willpartially be compensated by a simple result in section 4.5 called Barbalat's lemma.

TIME-VARYING POSITIVE DEFINITE FUNCTIONS AND DECRESCENTFUNCTIONS

When studying non-autonomous systems using Lyapunov's direct method, scalarfunctions with explicit time-dependence V(t, x) may have to be used, while in

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (120)

106 Advanced Stability Theory Chap. 4

autonomous system analysis time-invariant functions V(x) suffice. We now introducea simple definition of positive definiteness for such scalar functions.

Definition 4.7 A scalar time-varying function V(x, t) is locally positive definite ifV(0, 0 = 0 and there exists a time-invariant positive definite function Vo(x) such that

V f > ( 0 , V(x,t)>Vo(x) (4.7)

Thus, a time-variant function is locally positive definite if it dominates atime-invariant locally positive definite function. Globally positive definite functionscan be defined similarly.

Other related concepts can be defined in the same way, in a local or a globalsense. A function V(x, t) is negative definite if - V(x, t) is positive definite; V(x, t) ispositive semi-definite if V(x, t) dominates a time-invariant positive semi-definitefunction; V(x, t) is negative semi-definite if - V(x, t) is positive semi-definite.

In Lyapunov analysis of non-autonomous systems, the concept of decrescentfunctions is also necessary.

Definition 4.8 A scalar function V(x, t) is said to be decrescent if V(Q,t) = 0, and ifthere exists a time-invariant positive definite function V j(x) such that

V t > 0 , V(x,t)< V{(x)

In other words, a scalar function V(x, t) is decrescent if it is dominated by a time-invariant positive definite function.

Example 4.4: A simple example of a time-varying positive definite function is

V(x, r) = (1 +sin 2 r ) (x , 2 + x22)

because it dominates the function Vo(\) -xl 2 + x2

2. This function is also decrescent because it is

dominated by the function V,(x) = 2(x, 2 + x22)- D

Given a time-varying scalar function V(x, t), its derivative along a systemtrajectory is

^ = av + av/ . = 3v + a v / ( M (48 )

at dt dx dt 3x

LYAPUNOV THEOREM FOR NON-AUTONOMOUS SYSTEM STABILITY

The main Lyapunov stability results for non-autonomous systems can be summarizedby the following theorem.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (121)

Sect. 4.2 Lyapunov Analysis of Non-Autonomous Systems 107

Theorem 4.1 (Lyapunov theorem for non-autonomous systems)

Stability: If, in a ball BR around the equilibrium point 0, there exists a scalarfunction V(x, t) with continuous partial derivatives such that

1. V is positive definite

2. V is negative semi-definite

then the equilibrium point 0 is stable in the sense of Lyapunov.

Uniform stability and uniform asymptotic stability: If, furthermore,

3. V is decrescent

then the origin is uniformly stable. If condition 2 is strengthened by requiring thatV be negative definite, then the equilibrium point is uniformly asymptoticallystable.

Global uniform asymptotic stability: If the ball BR is replaced by the whole statespace, and condition 1, the strengthened condition 2, condition 3, and the condition

4. V(x, 0 is radially unbounded

are all satisfied, then the equilibrium point at 0 is globally uniformlyasymptotically stable.

Similarly to the case of autonomous systems, if, in a certain neighborhood ofthe equilibrium point, V is positive definite and V, its derivative along the systemtrajectories, is negative semi-definite, then V is called a Lyapunov function for thenon-autonomous system.

The proof of this important theorem, which we now detail, is rather technical.Hurried readers may skip it in a first reading, and go directly to Example 4.5.

In order to prove the above theorem, we first translate the definitions of positive definite

functions and decrescent functions in terms of the so-called class-K functions.

Definition 4.9 A continuous function a: R + —> R + is said to be of class K for to belong to class

K),if

• <x(0) = 0

• a(p) > 0 Vp > 0

• a is non-decreasing

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (122)

108 Advanced Stability Theory Chap. 4

The following lemma indicates the relation of positive definite and decrescent functions to

class K functions.

Lemma 4.1: A function V(x, t) is locally (or globally) positive definite if, and only if, there exists a

function a of class K such that V(0,0 = 0 and

V(x, t) > a(||x||) (4.9)

V (> 0 and V x e B g (or the whole state space).

A function V(x, t) is locally (or globally) decrescent if and only if there exists a class K

function (3 such that V(0, () = 0 and

V(x, r) < p(l(x|{) (4.10)

V t > 0 and V x e B^ (or in the whole state space).

Proof: Let us prove the positive definite function part first. Sufficiency is obvious from the

definition, because a(||x||) itself is a scalar time-invariant positive definite function. We now

consider necessity, i.e., assume that there exists a time-invariant positive function V0(x) such that

V(x, t) > Vo(x), and show that a function a of class K exists such that (4.9) holds. Let us define

a(p)= inf Vo(x) (4.11)p £ IMI £ R

Then, oc(0) = 0, a is continuous and non-decreasing. Because V0(x) is a continuous function and

non-zero except at 0, <x(p) > 0 for p > 0. Therefore, a is a class K function. Because of (4.11),

(4.9) is satisfied.

The second part of the lemma can be proven similarly, with the function P defined by

P(p) = sup V,(x) (4.12)

0 < ||x|| < p

where Vj(x) is the time-invariant positive function in Definition 4.8. D

Given the above lemma, we can now restate Theorem 4.1 as follows:

Theorem 4.1 Assume that, in a neighborhood of the equilibrium point 0, there exists a scalar

function V(x, () with continuous first order derivatives and a class-K function a such that, Vx ^t 0

2a. V(x, ( ) < 0

then the origin 0 is Lyapunov stable. If, furthermore, there is a scalar class-K function P such that

3. Hx,r)<P(||x||) J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (123)

Sect. 4.2 Lyapunov Analysis of Non-Autonomous Systems 109

then 0 is uniformly stable. If conditions 1 and 3 are satisfied and condition 2a is replaced by

condition 2b

with y being another class-K function, then 0 is uniformly asymptotically stable. If conditions 1, 2b

and 3 are satisfied in the whole state space, and

lim <x(||x|j)->°°x-> °°

then 0 is globally uniformly asymptotically stable.

Proof: We derive the three parts of the theorem in sequence.

Lyapunov stability: To establish Lyapunov stability, we must show that given R > 0, there exists

r > 0 such that (4.6) is satisfied. Because of conditions 1 and 2a,

<x(||x(0||) < V[x(f),'] * V[x(to), y Vt > to (4.13)

Because V is continuous in terms of x and V(0, to) = 0, we can find r such that

\\x(to)\\<r => V(x(to),to)<a(R)

This means that if ||x(/0)|| < r, then a(||x(/)j|) < a(R), and, accordingly, ||x(0|| <R , Vt> to.

Uniform stability and uniform asymptotic stability: From conditions 1 and 3,

<x(||xMII) < V(x(t), t) < P(||x(0||)

For any R > 0, there exists r(R) > 0 such that (}(r) < a(R) (Figure 4.1). Let the initial condition

x(t0) be chosen such that ||x(/0)|| < r. Then

a(R) > P(r) > V[x(to), to] > V[x(t), t] > a(||x(0ll)

This implies that

Vt>tn, \\x(t)\\<R

Uniform stability is asserted because r is independent of to .

In establishing uniform asymptotic stability, the basic idea is that if x does not converge to

the origin, then it can be shown that there is a positive number a such that - V [x(t), t] > a > 0 .

This implies that

V[x(t), t] - V[x(to), to]=\'vdt<-(t-t0)a

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (124)

110 Advanced Stability Theory Chap. 4

p (llxll)V ( x , t )

a(llxll)

R llxllFigure 4.1 : A positive definite and

decrescent function

and thus, that

0<V[x(t),t]<V[x(to),to]-(t-t0)a

which leads to a contradiction for large (. Let us now detail the proof.

Let | | x (g || < r, with r obtained as before. Let n be be any positive constant such that

0 <(i < || x ( g ||. We can find another positive constant 8(|X) such that (5(8)< a(n.). Define e = 7(5)

and set

Then, if ||x(r)|| > |i for all t in the period t0 < t < ty H t0 + T, we have

O'< a(u.) < V[x(r,),t!]< V[x(ro,g- f'VllxWID* < V[x{to),to)} - f

< V[x(f0), g - (f[ - g e < P(r) - Te = 0

This is a contradiction, and so there must exist l2 6 ['o.'il s » c n that ||x((2)|| < 5 . Thus, for all

t>,2.

a(||x(0ll) < V[x((),(] < V[x(r2),(2] < P(8) < a(ji)

As a result,

||x(0i|<M- Vr>( o +T>f 2

which shows uniform asymptotic stability.

Global uniform asymptotic stability: Since a ( ) is radially unbounded, R can be found such that

P(r) < a(R) for any ;\ In addition, /• can be made arbitrarily large. Hence, the origin x = 0 is

globally uniformly asymptotically stable. O

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (125)

Sect. 4.2 Lyapunov Analysis of Non-Autonomous Systems 111

Example 4.5: Global Asymptotic Stability

Consider the system defined by

i , M = - * , ( » ) - e - 2 ' j c 2 ( 0

x2(t) = x](t)-x2(t)

To determine the stability of the equilibrium point at 0, let us choose the following scalar function

This function is positive definite, because it dominates the time-invariant positive function

X|2 + x22 . I' ' s a ' s o decrescent, because it is dominated by the time-invariant positive definite

function xi 2 + 2x2

2 . Furthermore,

This shows that

Thus, V is negative definite, and therefore, the point 0 is globally asymptotically stable. IH

Stability results for non-autonomous systems can be less intuitive than those forautonomous systems, and therefore, particular care is required in applying the abovetheorem, as we now illustrate with two simple examples.

For autonomous systems, the origin is guaranteeed to be asymptotically stable ifV is positive definite and V is negative definite. Therefore, one might be tempted toconjecture that the same conditions are sufficient to guarantee the asymptotic stabilityof the system. However, this intuitively appealing statement is incorrect, asdemonstrated by the following counter-example.

Example 4.6: Importance of the decrescence condition

Let g(t) be a continuously-differentiable function which coincides with the function e~'12 except

around some peaks where it reaches the value 1. Specifically, g2(t) is shown in Figure 4.2. There

is a peak for each integer value of (. The width of the peak corresponding to abcissa n is assumed

to be smaller than (1/2)". The infinite integral of g2 thus satisfies

f g\r)dr < I e~rdr + T — = 2

and therefore, the scalar function

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (126)

112 Advanced Stability Theory

v2V(x, 0 = -

gl(t) Jo

is positive definite ( V(\, t) > x1).

Chap. 4

(4.14)

0 1 Figure 4 .2: The function g2(t)

Now consider the first-order differential equation

g{i)

If we choose V(x, t) in (4.14) as the Lyapunov function candidate, we easily find that

V=-X2

i.e., V is negative definite. Yet, the general solution of (4.15) is

(4.15)

and hence the origin is not asymptotically stable. •Since the positive definiteness of V and the negative semi-definiteness of V are

already sufficient to guarantee the Lyapunov stability of the origin, one may wonderwhat additional property the negative definiteness of V does provide. It can be shownthat, if V is negative definite, then an infinite sequence r,'s (/ = 1, 2,...) can be found

such that the corresponding state values \(tj) converge to zero as i —> °° (a result ofmostly theoretical interest).

Another illustration that care is required before jumping to conclusions involvesthe following interesting second-order dynamics

x + c{t)'x + kox = 0 (4.16)

which can represent a mass-spring-damper system (of mass 1), where c(t) >0 is atime-varying damping coefficient, and ko is a spring constant. Physical intuition maysuggest that the equilibrium point (0,0) is asymptotically stable as long as the damping

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (127)

Sect. 4.2 Lyapunov Analysis of Non-Autonomous Systems 113

c(t) remains larger than a strictly positive constant (implying constant dissipation ofenergy), as is the case for autonomous nonlinear mass-spring-damper systems.However, this is not necessarily true. Indeed, consider the system

x + (2+e')x + x = 0

One easily verifies that, for instance, with the initial condition JC(O) = 2, x(0) = - 1 , thesolution is x(t) - 1 + e~l, which tends to x = 1 instead! Here the damping increases sofast that the system gets "stuck" at x = 1.

Let us study the asymptotic stability of systems of the form of (4.16), using aLyapunov analysis.

Example 4.7: Asymptotic stability with time-varying damping

Lyapunov stability of the system (although not its asymptotic stability) can be easily established

using the mechanical energy of the system as a Lyapunov function. Let us now use a different

Lyapunov function to determine sufficient conditions for the asymptotic stability of the origin for

the system (4.16). Consider the following positive definite function

K ( x , o + x

where a is any positive constant smaller than -J ko , and

Kcan be easily computed as

V=la-c(t)]i2 + ^{c(!)-2ko}x2

Thus, if there exist positive numbers a and (5 such that

c(/)>a>0 c(f)<P<2/to

then V is negative definite. Assuming in addition that c(t) is upper bounded (guaranteeing the

decrescence of I7), the above conditions imply the asymptotic convergence of the system.

It can be shown [Rouche, et al., 1977] that, actually, the technical assumption that c(t) is

upper bounded is not necessary. Thus, for instance, the system

is asymptotically stable. [J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (128)

114 Advanced Stability Theory Chap. 4

4.2.2 Lyapunov Analysis of Linear Time-Varying Systems

None of the standard approaches for analyzing linear time-invariant systems (e.g.,eigenvalue determination) applies to linear time-varying systems. Thus, it is ofinterest to consider the application of Lyapunov's direct method for studying thestability of linear time-varying systems. Indeed, a number of results are available inthis regard. In addition, in view of the relation between the stability of a non-autonomous system and that of its (generally time-varying) linearization, to bediscussed in section 4.2.3, such results on linear time-varying systems can also be ofpractical importance for local stability analysis of nonlinear non-autonomous systems.

Consider linear time-varying systems of the form

x = A(r) x (4.17)

=-1 e2t~

0 -1xx

x2

Since LTI systems are asymptotically stable if their eigenvalues all have negative realparts, one might be tempted to conjecture that system (4.17) will be stable if at anytime t > 0, the eigenvalues of A(t) all have negative real parts. If this were indeed thecase, it would make the analysis of linear time-varying systems very easy. However,this conjecture is not true.

Consider for instance the system

(4.18)

The eigenvalues of the matrix A(/) are both equal to - 1 at all times. Yet, solving firstfor x2 and then substituting in the xx equation, one sees that the system verifies

x2 = x2(0) e~< xx+xx= x2(0) e'

and therefore is unstable, since x\ can be viewed as the output of a first-order filterwhose input x2(0) e' tends to infinity.

A simple result, however, is that the time-varying system (4.17) isasymptotically stable if the eigenvalues of the symmetric matrix A(0 + AT(t) (all ofwhich are real) remain strictly in the left-half complex plane

3\>0,\/i,\/t>0, \j(A(t) + Ar(0) < - *- (4.19)

This can be readily shown using the Lyapunov function V = x^x, since

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (129)

Sect. 4.2 Lyapunov Analysis of Non-Autonomous Systems 115

V = xTx + xTx = xr (A(f) + AT(t)) x<-XxTx = -lV

so that

V t > 0 , 0 < xrx = V(t) < V(0) e-Xt

and therefore x tends to zero exponentially.

Of course, the above result also applies in case the matrix A depends explicitlyon the state. It is also important to notice that the result provides a sufficient conditionfor asymptotic stability (some asymptotically stable systems do not verify (4.19), seeExercise 4.8).

A large number of more specialized theorems are available to determine thestability of classes of linear time-varying systems. We now give two examples of suchresults. The first result concerns "perturbed" linear systems, i.e., systems of the form(4.17), where the matrix A(t) is a sum of a constant stable matrix and some "small"time-varying matrix. The second result is a more technical property of systems suchthat the matrix A(t) maintains all its eigenvalues in the left-half plane and satisfiescertain smoothness conditions .

Perturbed linear systems

Consider a linear time-varying system of the form

x = (A, + A2(?)) x (4.20)

where the matrix Aj is constant and Hurwitz (i.e., has all its eigenvalues strictly in theleft-half plane), and the time-varying matrix A2(?) is such that

A2(0 -> 0 as t -> °°

and

oo

J II A2(0 II dt < °° (i.e, the integral exists and is finite)

Then the system (4.20) is globally exponentially stable.

Example 4.8: Consider the system

X2 = ~ X2 + 4 X3 2

.v3 = - (2 + sin I) jr3

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (130)

116 Advanced Stability Theory Chap. 4

Since JC3 tends to zero exponentially, so does JC 32 , and therefore, so does x2- Applying the above

result to the first equation, we conclude that the system is globally exponentially stable. CD

Sufficient smoothness conditions on the A(0 matrix

Consider the linear system (4.17), and assume that at any time t > 0, the eigenvalues ofA(t) all have negative real parts

3a>0,Vi,Vr>0,yA(0]<-a (4.21)

If, in addition, the matrix A(?) remains bounded, and

oo

f AT(t) A(t) dt < °° (i.e, the integral exists and is finite)!I

then the system is globally exponentially stable.

4.2.3 * The Linearization Method for Non-Autonomous Systems

I Lyapunov's linearization method can also be developed for non-autonomous systems.1 Let a non-autonomous system be described by (4.1) and 0 be an equilibrium point.

Assume that f is continuously differentiable with respect to x. Let us denoteI

(4.22)

Then for any fixed time t (i.e., regarding t as a parameter), a Taylor expansion of fleads to

If f can be well approximated by A(t)x for any time t, i.e.,

Urn sup h-°J *' = 0 Vf>0 (4.23)

||x|| -> 0 | |X| |

then the system

x = A(r)x (4.24)is said to be the linearization (or linear approximation) of the nonlinear non-autonomous system (4.1) around equilibrium point 0.

Note that

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (131)

Sect. 4.3 * Instability Theorems 117

• The Jacobian matrix A thus obtained from a non-autonomous nonlinearsystem is generally time-varying, contrary to what happens for autonomousnonlinear systems. But in some cases A is constant. For example, thenonlinear system x = -x + x2/t leads to the linearized system x = -x.

• Our later results require that the uniform convergence condition (4.23) besatisfied. Some non-autonomous systems may not satisfy this condition, andLyapunov's linearization method cannot be used for such systems. Forexample, (4.23) is not satisfied for the system x - -x + tx1.

Given a non-autonomous system satisfying condition (4.23), we can assert its(local) stability if its linear approximation is uniformly asymptotically stable, as statedin the following theorem:

Theorem 4.2 If the linearized system (with condition (4.23) satisfied) is uniformlyasymptotically stable, then the equilibrium point 0 of the original non-autonomoussystem is also uniformly asymptotically stable.

Note that the linearized time-varying system must be uniformly asymptoticallystable in order to use this theorem. If the linearized system is only asymptoticallystable, no conclusion can be drawn about the stability of the original nonlinear system.Counter-examples can easily verify this point.

Unlike Lyapunov's linearization method for autonomous systems, the abovetheorem does not relate the instability of the linearized time-varying system to that ofthe nonlinear system. There does exist a simple result which infers the instability of anon-autonomous system from that of its linear approximation, but it is applicable onlyto non-autonomous systems whose linear approximations are time-invariant.

Theorem 4.3 // the Jacobian matrix A(0 is constant, A(t) = Ao , and if (4.23) issatisfied, then the instability of the linearized system implies that of the original non-autonomous nonlinear system, i.e., (4.1) is unstable if one or more of the eigenvaluesof Ao has a positive real part.

*4.3 Instability Theorems

The preceding study of Lyapunov's direct method is concerned with providingsufficient conditions for stability of nonlinear systems. This section provides someinstability theorems based on Lyapunov's direct method.

Note that, for autonomous systems, one might think that the conclusive resultsprovided by Lyapunov's linearization method are sufficient for the study of instability.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (132)

118 Advanced Stability Theory Chap. 4 |J

However, in some cases, instability theorems based on the direct method may beadvantageous. Indeed, if the linearization method fails (i.e., if the linearized system ismarginally stable), these theorems may be used to determine the instability of thenonlinear system. Another advantage may be convenience, since the theorems do notrequire linearization of the system equations.

We state the theorems directly in a non-autonomous setting. For autonomoussystems, the conditions simplify in a straightforward fashion.

Theorem 4.4 (First instability theorem) If, in a certain neighborhood Q of theorigin, there exists a continuously differentiable, decrescent scalar function V(x, t)such that

• V(0, 0 = 0 V t > t0

• V(x, t0) can assume strictly positive values arbitrarily close to the origin

• V(x, 0 is positive definite (locally in Q.)

then the equilibrium point 0 at time t0 is unstable.

Note that the second condition is weaker than requiring the positive definitenessof V. For example, the function V(x) = jti2 - x2

2 is obviously not positive definite,but it can assume positive values arbitrarily near the origin (V(x) = Xj2 along the line

Example 4.9: Consider the system

x, = 2 x 2 + x , ( x , 2 + 2x24) (4.26)

x2 = - 2 x , + .v2(x,2 + x24) (4.27)

Linearization of this system yields Jt, = 2x2 and x2 = - 2xj. The eigenvalues of this system are

+ 2) and - 2y", indicating the inability of Lyapunov's linearization method for this system.

However, if we take

2 x\ X1

its derivative is

Because of the positive definiteness of V and V, the above theorem indicates the instability of the

system.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (133)

Sect. 4.3 Instability Theorems 119

Theorem 4.5 (Second instability theorem) If, in a certain neighborhood Q. of theorigin, there exists a continuously differentiable, decrescent scalar function V(x, t)satisfying

• V(0, t0) = 0 and V(x, t0) can assume strictly positive values arbitrarilyclose to the origin

• V(x, t) - X V(x, t) > 0 V t > t0 V x e Q

with X being a strictly positive constant, then the equilibrium point 0 at time t0 isunstable.

Example 4.10: Consider the system described by

X] = X| + 3x 2 s in 2 x 2 + 5xjX22sin2X| (4.28)

x2 = 3xj sin2x2 + x2 - 5x 12 x 2 cos 2 x 1 (4.29)

Let us consider the function K(x) = ( l / 2 ) ( X j 2 —x22) , which was shown earlier to assume

positive values arbitrarily near the origin. Its derivative is

V = x , 2 - x 22 + 5 x , 2 x 2

2 = 2V + 5 x , 2 x 22

Thus, the second instability theorem shows that the equilibrium point at the origin is unstable. Of

course, in this particular case, the instability could be predicted more easily by the linearization

method. CD

In order to apply the above two theorems, V is required to satisfy certainconditions at all points in the neighborhood £2. The following theorem (Cetaev'stheorem) replaces theses conditions by a boundary condition on a subregion in O.

Theorem 4.6 (Third instability theorem) Let Q. be a neighborhood of the origin.If there exists a scalar function V(x, 0 with continuous first partial derivatives,decrescent in Q, and a region O/ in O, such that

• K(x, t) and V(\, t) are positive definite in Qj

• The origin is a boundary point ofQ./

• At the boundary points ofQ.] within Q, V(x, t) = Ofor all t>t0

then the equilibrium point 0 at time to is unstable.

The geometrical meaning of this theorem can be seen from Figure 4.3. Let usillustrate the use of this theorem on a simple example.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (134)

120 Advanced Stability Theory Chap. 4

n1 (V>oV>0)

Figure 4.3 : Geometrical interpretation of

the third instability theorem

Example 4.11: Consider the system

The linearization of this system leads to a pole at the origin and a pole at - 1. Therefore,

Lyapunov's linearization method cannot be used to determine the stability of this nonlinear

system. Now let us take the function V = x, - x22/2. Its derivative is

Examining V and V and using Cetaev's theorem, one can show the instability of the origin. •4.4 * Existence of Lyapunov Functions

In the previous Lyapunov theorems, the existence of Lyapunov functions is alwaysassumed, and the objective is to deduce the stability properties of the systems from theproperties of the Lyapunov functions. In view of the common difficulty in findingLyapunov functions, one may naturally wonder whether Lyapunov functions alwaysexist for stable systems. A number of interesting results concerning the existence ofLyapunov functions, called converse Lyapunov theorems, have been obtained in thisregard. For many years, these theorems were thought to be of no practical "valuebecause, like the previously described theorems, they do not tell us how to generateLyapunov functions for a system to be analyzed, but only represent comfortingreassurances in the search for Lyapunov functions. In the past few years, however,there has been a resurgence of interest in these results. The reason is that a subsystemof a nonlinear system may be known to possess some stability properties, and theconverse theorems allow us to construct a Lyapunov function for the subsystem,which may subsequently lead to the generation of a Lyapunov function for the wholesystem. In particular, the converse theorems can be used in connection with stability

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (135)

Sect. 4.4 * Existence of Lyapunov Functions 121

analysis of feedback linearizable systems and robustness analysis of adaptive controlsystems.

THE CONVERSE THEOREMS

There exists a converse theorem for essentially every Lyapunov stability theorem(stability, uniform stability, asymptotic stability, uniform asymptotic stability, globaluniform asymptotic stability and instability). We now present three of the conversetheorems.

Theorem 4.7 (stability) If the origin of (4.1) is stable, there exists a positivedefinite function V(x, t) with a non-positive derivative.

This theorem indicates the existence of a Lyapunov function for every stable system.

Theorem 4.8 (uniform asymptotic stability) If the equilibrium point at the originis uniformly asymptotically stable, there exists a positive definite and decrescentfunction V(x, t) with a negative definite derivative.

This theorem is theoretically important because it will later be useful in establishingrobustness of uniform asymptotic stability to persistent disturbance.

The next theorem on exponential stability has more practical value than thesecond theorem, because its use may allow us to explicitly estimate convergence ratesin some nonlinear systems.

Theorem 4.9 (exponential stability) If the vector function f(x, t) in (4.1) hascontinuous and bounded first partial derivatives with respect to x and t,for all x in aball B,. and for all t>0, then the equilibrium point at the origin is exponentially stableif, and only if, there exists a function V(x, t) and strictly positive constants o.j, OLj, (Xj,a4 such that V x e B,., V t > 0

cqIMI2 < V(x,t) < a2 | |x||2 (4.30)

V < - a 3 | | x | | 2 (4.31)

l l ^ l l ^ oc4||x|| (4.32)d x

The proofs of the converse theorems typically assume that the solution of thesystem is available, and then construct a Lyapunov function based on the assumedsolution [Hahn, 1967]. Proof of Theorem 4.9 can be found in [Bodson, 1986; SastryandBodson, 1989].

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (136)

122 Advanced Stability Theory Chap. 4

4.5 Lyapunov-Like Analysis Using Barbalat's Lemma

For autonomous systems, the invariant set theorems are powerful tools to studystability, because they allow asymptotic stability conclusions to be drawn even whenV is only negative .se/Mj-definite. However, the invariant set theorems are notapplicable to non-autonomous systems. Therefore, asymptotic stability analysis ofnon-autonomous systems is generally much harder than that of autonomous systems,since it is usually very difficult .to find Lyapunov functions with a negative definitederivative. An important and simple result which partially remedies this situation isBarbalat's lemma. Barbalat's lemma is a purely mathematical result concerning theasymptotic properties of functions and their derivatives. When properly used fordynamic systems, particularly non-autonomous systems, it may lead to the satisfactorysolution of many asymptotic stability problems.

4.5.1 Asymptotic Properties of Functions and Their Derivatives

Before discussing Barbalat's lemma itself, let us clarify a few points concerning theasymptotic properties of functions and their derivatives. Given a differentiablefunction/of time t, the following three facts are important to keep in mind.

• /—> 0 ?t> /converges

The fact that/(0 —> 0 does not imply that/(f) has a limit as t —> °° .

Geometrically, a diminishing derivative means flatter and flatter slopes.However, it does not necessarily imply that the function approaches a limit.Consider, for instance, the rather benign function/(0 = sin(log t). While

cos(loe t) n= — v B ' -> 0 as t-*°°

t

the function/(0 keeps oscillating (slower and slower). The function fit)may even be unbounded, as with/(f) = -\[7 sin(log f). Note that functions ofthe form log t, sin t, e01, and combinations thereof, are quite easy to find indynamic system responses.

/converges &> /—»0

The fact that/(r) has a finite limit as t —> °° does not imply that/(f) -> 0.

For instance, while the function/(f) = e~' sin(e2') tends to zero, its derivative/ i s unbounded. Note that this is not linked to the frequent sign changes of the

'i

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (137)

Sect. 4.5 Lyapunov-Like Analysis Using Barbalat's Lemma 123

function. Indeed, with/(?) = e~' sin2(e2') > 0 , / is still unbounded.

• If / i s lower bounded and decreasing ( / < 0 ) , then it converges to alimit.

This is a standard result in calculus. However, it does not say whether theslope of the curve will diminish or not.

4.5.2 Barbalat's Lemma

Now, given that a function tends towards a finite limit, what additional requirementcan guarantee that its derivative actually converges to zero? Barbalat's lemmaindicates that the derivative itself should have some smoothness. More precisely, wehave

Lemma 4.2 (Barbalat) If the differentiable function f{i) has a finite limit ast —> °°, and if f is uniformly continuous, then f(t) —> 0 as t —> °° .

Before proving this result, let us define what we mean by uniform continuity.Recall that a function g(t) is continuous on [0, °°) if

Vff>0, V R>0,3T)(R,tl)>0, V?>0, U-r , l<T | => \g(t)-g(tt)l<R

A function g is said to be uniformly continuous on [0, °°) if

VR>0,3r\(R)>0;Vtl>0,Vt>0, \t-tl\<r\ => lg(0-g('i)l <R

In other words, g is uniformly continuous if one can always find an T) which does notdepend on the specific point f j - and in particular, such that T| does not shrink as

?1 —¥ °° , as shall be important when proving the lemma. Note that t and fj play asymmetric role in the definition of uniform continuity.

Uniform continuity of a function is often awkward to assert directly from theabove definition. A more convenient approach is to examine the function's derivative.Indeed, a very simple sufficient condition for a differentiable function to be uniformlycontinuous is that its derivative be bounded. This can be easily seen from the finitedifference theorem

V f , V f [ , 3 t2 (between rand t^ such that g(t) - g(t^) = g(t2) (t- t{)

and therefore, if Rl>0 is an upper bound on the function l£l , one can always useT| = R/R j independently of /j to verify the definition of uniform continuity.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (138)

124 Advanced Stability Theory Chap. 4

Let us now prove Barbalat's lemma, by contradiction.

Proof of Barbalat's lemma: Assume that f(t) does not approach zero as (-> °°. Then

3 e0 > 0, V T > 0 , 3 t > T, 1/(0 I > eo . Therefore, we can get an infinite sequence of /,'s (such

that (,- —> °° as ( —> °o) such that l/(f,-) I Ro • Since/(0 is assumed to be uniformly continuous,

3 r| > 0, such that for any / and ( satisfying \t —t \<v\

\hn -/('")i <~

This implies that for any t within the ^-neighborhood of ti (i.e., such that 1/ - (,1 < r\)

1/(0 I > y

Hence, for all r ;,

r',+y\ • r';+Tl • e oif f(t)dt\ = f ' 1/1(0* ^ ^

' - n ' - r i lif f

where the left equality comes from the fact that / keeps a constant sign over the integration

interval, due to the continuity of/and the bound 1/(0 I > eo/2 > 0 .

This result would contradict the known fact that the integral [ f(r)dr has a limit (equal to

/(oo) _/(0)) as t -> oo . •

Given the simple sufficient condition for uniform continuity mentioned earlier,an immediate and practical corollary of Barbalat's lemma can be stated as follows: (/the differentiable function f (t) has a finite limit as f-> °°, and is such that f existsand is bounded, thenf{t) —> 0 as t —> °° .

The following example illustrates how to assert the uniform continuity ofsignals in control systems.

Example 4.12: Consider a strictly stable linear system whose input is bounded. Then the system

output is uniformly continuous.

Indeed, write the system in the standard form

x = A x + B u

y = Cx

Since u is bounded and the linear system is strictly stable, thus the state x is bounded. This in turn

implies from the first equation that x is bounded, and therefore from the second equation that

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (139)

Sect. 4.5 Lyapunov-Like Analysis Using Barbalat's Lemma 125

y = C x is bounded. Thus the system output y is uniformly continuous. D

USING BARBALAT'S LEMMA FOR STABILITY ANALYSIS

To apply Barbalat's lemma to the analysis of dynamic systems, one typically uses thefollowing immediate corollary, which looks very much like an invariant set theorem inLyapunov analysis:

Lemma 4.3 ("Lyapunov-Like Lemma") If a scalar function V(x, t) satisfies thefollowing conditions

• V(x, t) is lower bounded

• V^x, t) is negative semi-definite

• V^x, t) is uniformly continuous in time

then V{x, t) -» 0 as t ->°« .

Indeed, V then approaches a finite limiting value VQO, such that Voo 2 V (x(0), 0) (thisdoes not require uniform continuity). The above lemma then follows from Barbalat'slemma.

To illustrate this procedure, let us consider the asymptotic stability analysis of asimple adaptive control system.

Example 4.13: As we shall detail in chapter 8, the closed-loop error dynamics of an adaptivecontrol system for a first-order plant with one unknown parameter is

e = -e + 6w(t)

Q = -ew(t)

where e and 8 are the two states of the closed-loop dynamics, representing tracking error and

parameter error, and w{t) is a bounded continuous function (in the general case, the dynamics has

a similar form but with e, 0, and w(t) replaced by vector quantities). Let us analyze the

asymptotic properties of this system.

Consider the lower bounded function

Its derivative is

This implies that V(t) < V(0), and therefore, that e and 9 are bounded. But the invariant set

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (140)

126 Advanced Stability Theory Chap. 4

theorems cannot be used to conclude the convergence of e, because the dynamics is non-

autonomous.

To use Barbalat's lemma, let us check the uniform continuity of V. The derivative of V is

- 7

This shows that V is bounded, since w is bounded by hypothesis, and e and 0 were shown above ~

to be bounded. Hence, V is uniformly continuous. Application of Barbalat's lemma then

indicates that e —> 0 as t —> °° .

Note that, although e converges to zero, the system is not asymptotically stable, because 8 is

only guaranteed to be bounded. Q

The analysis in the above example is quite similar to a Lyapunov analysis basedon invariant set theorems. Such an analysis based on Barbalat's lemma shall be calleda Lyapunov-like analysis. It presents two subtle but important differences withLyapunov analysis, however. The first is that the function V can simply be a lowerbounded function of x and t instead of a positive definite function. The seconddifference is that the derivative V must be shown to be uniformly continuous, inaddition to being negative or zero. This is typically done by proving that V is bounded.Of course, in using the Lyapunov-like lemma for stability analysis, the primarydifficulty is still the proper choice of the scalar function V.

4.6 Positive Linear Systems

In the analysis and design of nonlinear systems, it is often possible and useful todecompose the system into a linear subsystem and a nonlinear subsystem. If thetransfer function (or transfer matrix) of the linear subsystem is so-called positive real,then it has important properties which may lead to the generation of a Lyapunovfunction for the whole system. In this section, we study linear systems with positivereal transfer functions or transfer matrices, and their properties. Such systems, calledpositive linear systems, play a central role in the analysis and design of manynonlinear control problems, as will be seen later in the book.

4.6.1 PR and SPR Transfer Functions

We consider rational transfer functions of nth-order single-input single-output linearsystems, represented in the form

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (141)

Sect. 4.6 Positive Linear Systems 127

b nm + b i n m ~ l + + b,, , "mI um-\" T ••• T "oKp) =

The coefficients of the numerator and denominator polynomials are assumed to be realnumbers and n>m. The difference n-m between the order of the denominator andthat of the numerator is called the relative degree of the system.

Definition 4.10 A transfer function h(p) is positive real if

Re[h(p)]>0 for all Re[p]>0 (433)

It is strictly positive real if h(p-e) is positive real for some £ > 0 .

Condition (4.33), called the positive real condition, means that h(p) always has apositive (or zero) real part when p has positive (or zero) real part. Geometrically, itmeans that the rational function h(p) maps every point in the closed right half (i.e.,including the imaginary axis) of the complex plane into the closed right half of theh(p) plane. The concept of positive real functions originally arose in the context ofcircuit theory, where the transfer function of a passive network (passive in the sensethat no energy is generated in the network, e.g., a network consisting of onlyinductors, resistors, and capacitors) is rational and positive real. In section 4.7, weshall reconcile the PR concept with passivity.

Example 4.14: A strictly positive real function

Consider the rational function

Ih(p) =

p + X

which is the transfer function of a first-order system, with X > 0. Corresponding to the complex

variable/? =

Obviously, Re[/i(p)] > 0 if a > 0. Thus, h(p) is a positive real function. In fact, one can easily

see that h(p) is strictly positive real, for example by choosing e = X/2 in Definition 4.9. D

For higher-order transfer functions, it is often difficult to use the definitiondirectly in order to test the positive realness condition, because this involves checkingthe positivity condition over the entire right-half of the complex plane. The followingtheorem can simplify the algebraic complexity.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (142)

128 Advanced Stability Theory Chap. 4

Theorem 4.10 A transfer function h(p) is strictly positive real (SPR) if and only if

i) h(p) is a strictly stable transfer function

ii) the real part ofh(p) is strictly positive along thej(i> axis, i.e.,

V co > 0 Re[/j(/to)] > 0 (4.34)

The proof of this theorem is presented in the next section, in connection with the so-called passive systems.

The above theorem implies simple necessary conditions for asserting whether agiven transfer function h(p) is SPR:

• h(p) is strictly stable

• The Nyquist plot of h(f<x>) lies entirely in the right half complex plane.Equivalently, the phase shift of the system in response to sinusoidal inputs isalways less than 90°

• h(p) has relative degree 0 or 1

• h{p) is strictly minimum-phase {i.e., all its zeros are strictly in the left-half plane)

The first and second conditions are immediate from the theorem. The last twoconditions can be derived from the second condition simply by recalling the procedurefor constructing Bode or Nyquist frequency response plots (systems with relativedegree larger than 1 and non-minimum phase systems have phase shifts larger than90° at high frequencies, or, equivalently have parts of the Nyquist plot lying in theleft-half plane).

Example 4.15: SPR and non-SPR transfer functions

Consider the following systems

p + a p + b

h2(p) =p2-p+\

p1+ap+b

p2 + p + 1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (143)

Sect. 4.6 Positive Linear Systems 129

The transfer functions h^ h2 , and ft3 are not SPR, because hx is non-minimum phase, h2 is

unstable, and /i3 has relative degree larger than 1.

Is the (strictly stable, minimum-phase, and of relative degree 1) function h4 actually SPR?

We have

h (ico) = -/'t0+ ' = [/<*>+ l ] [ - (Q 2 -ya )+ 1]-co 2 +yco+l [ l - c o 2 ] 2 + co2

(where the second equality is obtained by multiplying numerator and denominator by the complex

conjugate of the denominator) and thus

„ . , . . , , -tf>2+ 1 + co2 1Re[ h4(jm) ] = [ 1 - C O 2 ] 2 + CO2 [1 - C l ) 2 ] 2 + CO2

which shows that h4 is SPR (since it is also strictly stable). Of course, condition (4.34) can also

be checked directly on a computer. Q

The basic difference between PR and SPR transfer functions is that PR transferfunctions may tolerate poles on the jo) axis, while SPR functions cannot.

Example 4.16: Consider the transfer function of an integrator,

h(p) = -P

Its value corresponding to p = a +j(£> is

One easily sees from Definition 4.9 that h(p) is PR but not SPR. •

More precisely, we have the following result, which complements Theorem4.10.

Theorem 4.11 A transfer function h(p) is positive real if, and only if,

i) h(p) is a stable transfer function

(il) The poles of h(p) on the ja axis are simple (i.e., distinct) and theassociated residues are real and non-negative

Hi) Re[h(j(i>)] > 0 for any co > 0 such that ;co is not a pole ofh(p)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (144)

130 Advanced Stability Theory

4.6.2 The Kalman-Yakubovich Lemma

Chap. 4

If a transfer function of a system is SPR, there is an important mathematical propertyassociated with its state-space representation, which is summarized in the celebratedKalman-Yakubovich (KY) lemma.

Lemma 4.4 (Kalman-Yakubovich)system

x=Ax + bw

y = cTx

Consider a controllable linear time-invariant

The transfer function

Mp) = cT (4.35)

is strictly positive real if, and only if, there exist positive definite matrices P and Qsuch that

ArP

Pb = c

= - Q (4.36a)

(4.36b)

The proof of this lemma is presented in section 4.7 in connection with passivity inlinear systems. Beyond its mathematical statement, which shall be extensively used inchapter 8 (Adaptive Control), the KY lemma has important physical interpretationsand uses in generating Lyapunov functions, as discussed in section 4.7.

The KY lemma can be easily extended to PR systems. For such systems, it canbe shown that there exist a positive definite matrix P and a positive sera-definitematrix Q such that (4.36a) and (4.36b) are verified. The usefulness of this result is thatit is applicable to transfer functions containing a pure integrator ( l/p in thefrequency-domain), of which we shall see many in chapter 8 when we study adaptivecontroller design. The Kalman-Yakubovich lemma is also referred to as the positivereal lemma.

In the KY lemma, the involved system is required to be asymptotically stableand completely controllable. A modified version of the KY lemma, relaxing thecontrollability condition, can be stated as follows:

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (145)

Sect. 4.6 Positive Linear Systems 131

Lemma 4.5 (Meyer-Kalman-Yakubovich) Given a scalar y> 0, vectors b and c,an asymptotically stable matrix A, and a symmetric positive definite matrix L, ;/ thetransfer function

//(p) = 1 + cT[pl-A]-lb

is SPR, then there exist a scalar e > 0, a vector q, and a symmetric positive definitematrix P such that

ATP + PA = - q q 7 - e L

Pb = c [~

This lemma is different from Lemma 4.4 in two aspects. First, the involvedsystem now has the output equation

y = cTx + %u

Second, the system is only required to be stabilizable (but not necessarilycontrollable).

4.6.3 Positive Real Transfer Matrices

The concept of positive real transfer function can be generalized to rational positivereal matrices. Such generalization is useful for the analysis and design of multi-input-multi-output nonlinear control systems.

Definition 4.11 An mxm transfer matrix H(p) is called PR if

H(p) has elements which are analytic for Re(p) > 0;

H(p) + UT(p*) is positive semi-definite for Re(p) > 0.

where the asterisk * denotes the complex conjugate transpose. H(p) is SPR ifH(p - e) is PR for some e > 0.

The Kalman-Yakubovich lemma and the Meyer-Kalman-Yakubovich lemma can beeasily extended to positive real transfer matrices.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (146)

132 Advanced Stability Theory Chap. 4

4.7 The Passivity Formalism

As we saw earlier, Lyapunov functions are generalizations of the notion of energy in adynamic system. Thus, intuitively, we expect Lyapunov functions to be "additive",i.e., Lyapunov functions for combinations of systems to be derived by simply addingthe Lyapunov functions describing the subsystems. Passivity theory formalizes thisintuition, and derives simple rules to describe combinations of subsystems or "blocks"expressed in a Lyapunov-like formalism. It also represents an approach toconstructing Lyapunov functions or Lyapunov-like functions for feedback controlpurposes.

As a motivation, recall first that the dynamics of state-determined physicalsystems, whether linear or nonlinear, satisfy energy-conservation equations of theform

— [Stored Energy] = [External Power Input] + [Internal Power Generation]dt

These equations actually form the basis of modeling techniques such as bond-graphs.The external power input term can be represented as the scalar product y^u of an input("effort" or "flow") u, and a output ("flow" or "effort") y.

In the following, we shall more generally consider systems which verifyequations of the form

y1Tu1 - gl{t) (4.37)

where Vj(f) and gj(O are scalar functions of time, Uj is the system input, and yt is itsoutput. Note that, from a mathematical point of view, the above form is quite general(given an arbitrary system, of input U)(?) and output y\(t), we can let, for instance,g](0 = 0 and V((r) = [' y^(r) Uj(r) dr). It is the physical or "Lyapunov-like"

properties that V\{t) and gj(f) may have, and how they are transmitted throughcombinations with similar systems, that we shall be particularly interested in.

4.7.1 Block Combinations

Assume that we couple a system in the form (4.37), or power form, to one verifyingthe similar equation

V2(t) = y2T u2 - g2(t)

in a feedback configuration, namely u2 = yj and U[ = - y2 (Figure 4.4), assuming of

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (147)

Sect. 4.7 The Passivity Formalism 133

course that the vectors u,- and y.• are all of the same dimension. We then have

dtVx(t) + V2(t) ] = - (4.38)

,, 82 Figure 4.4 : Two blocks of the form(4.37), in a feedback configuration

Let us assume that the function V\ + V2 is lower bounded (e.g., is positive). Then,using the same reasoning as in section 4.5, we have

• If V t > 0 , g\(i) + g2(t) ^ 0 , then the function Vj + V2 is upper bounded,and

< 00

If in addition, the function gj + g2 is uniformly continuous, then

->0 as t->°°.

• In particular, if gj(r) and g2(t) are both non-negative and uniformly

continuous, then they both tend to zero as t —> °°

Note that an explicit expression of Vl + V2 is not needed in the above results. Moregenerally, without any assumption on the sign of Vj + V2 or g\ + g2 , we can statethat

• If Vj + V2 has a finite limit as t —> °° , and if gj + g2 is uniformly

continuous, then [ gy(t) + g2(t) ] —> 0 as t —> °°.

A system verifying an equation of the form (4.37) with V\ lower bounded andg j > 0 is said to be passive (or to be a passive mapping between ii[ and yj).Furthermore, a passive system is said to be dissipative if

=> gl(t)dt>0

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (148)

134 Advanced Stability Theory Chap. 4

Example 4.17: The nonlinear mass-spring-damper system

m x + x^ x^ + x^ — F

represents a dissipative mapping from external force F to velocity i , since

i (Im i2+I;t8) = xF - x2k*

Of course, here Vj is simply the total (kinetic plus potential) energy stored in the system, and |

gj is the dissipated power. Q

Example 4.18: Consider the system (Figure 4.5)

x + X(t) x=u (4.39)

(4.39)

FILTER

h(x)i

NONLINEARITY

Figure 4.5 : A passive single-input single-output system

where the function h is of the same sign as its argument, although not necessarily continuous, and

X(t) > 0. The mapping w —> y is passive, since

4 \Xh(\)a\ = h(x)x = y u - X(t) h(x)xdtJo

with \xh(E,)dE, > 0 and X(t)h(x)x > 0 for all x. The mapping is actually dissipative if X(t) is

not identically zero.

Of course, the function X(t) may actually be of the form X[x{t)]. For instance, the system

y = x — sin 2x

is a dissipative mapping from u to y.

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (149)

Sect. 4.7 The Passivity Formalism 135

A particularly convenient aspect of formalizing the construction of Lyapunov-like functions as above, is that parallel and feedback combinations of systems in thepower form are still in the power form. Indeed, it is straightforward to verify that, forboth the parallel combination and the feedback combination (Figure 4.6), one has

y ru = y ^ u , + y 2r u 2

Namely, for the parallel combination

yTu = (y, + y2)Tu= y\Tu + y2

Tu = yir«i + y2r"2

and for the feedback combination

yTu = y1T(ul+y2) = y , ru, y{

Ty2 = u2T y2

Incidentally, this result is a particular case of what is known in circuit theory asTellegen's power conservation theorem. Thus, we have, for the overall system

V=V, = 81+82

By induction, any combination of feedback and/or parallel combinations ofsystems in the power form can also be described in the power form, with thecorresponding V and g simply being equal to the sum of the individual Vj and gj.

8 =

y=yi

" 2 = yi

Figure 4.6 : Parallel and feedback combinations

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (150)

136 Advanced Stability Theory Chap. 4

The power of this simple result is, of course, that it does not require that thesubsystems be linear.

Note that, assuming that V is lower bounded, the overall system can be passivewhile some of its components may be "active" ( g,- < 0 ): for the system to be passive,the sum of the g± has to be positive, i.e., there should be more overall "powerdissipation" in the system than there is "power generation". Also, note that thepassivity of a block is preserved if its input or its output is multiplied by a strictlypositive constant (an input gain or an output gain), since this simply multiplies theassociated V(- and gt by the same constant. Thus we have, more generally

where a, is the product of the input gain and the output gain for block i.

Example 4.19: Consider again the adaptive control system of Example 4.13. The fact that

e2 = e'e - e 6 w(t) - e 2

2 d;

can be interpreted as stating that the mapping 6 w(t) -4 c is dissipative. Furthermore, using the

parameter adaptation law

6 = - e w(t) (4.40)

then corresponds to inserting a passive feedback block between e and — 8 w(t), since

l i e 2 =-e>v(oe •2 d(

Note that, for a passive system, we may always assume that g = 0 , simply byadding [' g(r) dr to the original V. Hence, the definition of passivity is often written

as

3 a > - t > o j V ( > 0 , [' yT(r) u(r) dr > a (4.41)o

which simply says that there exists a lower bounded V such that g = 0 .

Also, note that the power form is expressed in terms of the dot-product y^u.Therefore, if ua and ya are other choices of inputs and outputs for the system such thatya

T ua = y^u at all times, then they satisfy the same passivity properties as u and y.For instance, if the mapping u —> y is passive, so is the mapping Au —» A~^y,where the matrix A is any (perhaps time-varying) invertible matrix. In particular,

j

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (151)

Sect. 4.7 The Passivity Formalism 137

passivity is conserved through orthogonal transformations ( A AT = I) . Furthermore,note that the dimension of the vectors ua and ya is not necessarily the same as thatof the vectors u and y .

4.7.2 Passivity in Linear Systems

An important practical feature of the passivity formulation is that it is easy tocharacterize passive linear systems. This allows linear blocks to be straightforwardlyincorporated or added in a nonlinear control problem formulated in terms of passivemappings.

As we shall now show, a strictly stable linear SISO system is passive if, andonly if,

V co > 0 , Re[h(j(o)] > 0 (4.42)

where h is the transfer function of the system (which we shall assume to be rational)and Re refers to the real part of a complex number. Geometrically, condition (4.42)can also be written as (Figure 4.7)

V co > 0 , | Arg h(j(0) | < ^ (4.43)

Thus, we see that a strictly stable linear SISO system is passive if, and only if itsphase shift in response to a sinusoidal input is always less than (or equal to) 90°.

Re Figure 4.7 : Geometric interpretation of

the passivity condition for a linear SISO

system

Proof of condition (4.42): The proof is based on Parseval's theorem, which relates the time-

domain and frequency-domain expressions of a signal's squared norm or "energy", as well as

those of the correlation of two signals.

Consider a strictly stable linear SISO system, of transfer function y(p)/u(p) = h(p), initially

at rest (y = 0) at t = 0. Let us apply to this system an arbitrary control input between / = 0 and

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (152)

138 Advanced Stability Theory Chap. 4

some positive time (, and no input afterwards. Recalling expression (4.41) of passivity, we

compute

I y(r) u(r) dr = \ y{r) u(r) dr = — | y(ja>) H*O'CO) daJ

o -Loo 271 -LOO

where the first equality comes from the fact that u is zero after time / (and, of course, both u and y

are zero for negative time), and the second equality comes from Parseval's theorem, with the

superscript referring to complex conjugation. Since y(jin) ~ h(jfo) «(/co), we thus have

; , oo| y{r) u(r) dr = — [ h(j(i>) |«O'co)|2 daJ /, 271 J _ no

Now since h is the transfer function of a real system, its coefficients are real, and thus

h(-j(H) = [h(ja)T . Hence,

f y(r)u(r)dr = - f ReWja))} \u{ja)\2 da (4.44)

Given expression (4.41) of passivity, equation (4.44) shows that (4.42) is a sufficient

condition for the system to be passive. Indeed, taking an arbitrary input u, the integral

f y(r) u(r) dr does not depend on the values of u at times later than ( (so that our earlier

assumption that u is zero after time (is not restrictive).

Equation (4.44) also shows that (4.42) is a necessary condition for the system to be passive.

Indeed, if (4.44) was not verified, then there would be a finite interval in co over which

Re[/i(yw)] < 0, because h is continuous in co. The integral could then be made arbitrarily negative

by choosing |«(y'to)| large enough over this finite interval. D

Note that we have assumed that h(p) is strictly stable, so as to guarantee theexistence of the frequency-domain integrals in the above proof. Actually, usingstandard results in complex variable analysis, the proof can be extended easily to thecase where h(p) has perhaps some poles on the yd) axis, provided that these poles besimple (i.e., distinct) and that the associated residues be non-negative. As discussedearlier in section 4.6, systems belonging to this more general class and verifyingcondition (4.42) are called positive real (PR) systems. Thus, a linear single-inputsystem is passive if (and only if) it is positive real.

Condition (4.42) can also be formally stated as saying that the Nyquist plot of his in the right half-plane. Similarly, if the Nyquist plot of a strictly stable (or PR)linear system of transfer function h is strictly in the right half-plane (except perhapsfor co = °° ), i.e, if

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (153)

Sect. 4.7 The Passivity Formalism 139

V co > 0 , Re[/*(/0))] > 0 (4.45)

then the system is actually dissipative. As discussed in section 4.6, strictly stable linearsystems verifying (4.45) are called strictly positive real (SPR) systems.

It can also be shown that, more generally, a strictly stable linear MIMO systemis passive if, and only if

V co > 0 , HO'co) + Hr(-;co) > 0

where H is the transfer matrix of the system. It is dissipative if

V co > 0 , H(yco) + Hr(-;co) > 0

THE KALMAN-YAKUBOVICH LEMMA

For linear systems, the closeness of the concepts of stability and passivity can beunderstood easily by considering the Lyapunov equations associated with the systems,as we now show. The discussion also provides a more intuitive perspective on the KYlemma of section 4.6.2, in the light of the passivity formalism.

Recall from our discussion of Lyapunov functions for linear systems (section3.5.1) that, for any strictly stable linear system of the form x = A x , one has

V Q symmetric p.d. , 3 P symmetric p.d. , such that AT P + P A = - Q (4.46)

an algebraic matrix equation which we referred to as the Lyapunov equation for thelinear system. Letting

= ~xTPx

yields

V=--xTQx

Consider now a linear system, strictly stable in open-loop, in the standard form

x = A x + B u y = Cx

The Lyapunov equation (4.46) is verified for this system, since it is only related to thesystem's stability, as characterized by the matrix A, and is independent of the inputand output matrices B and C. Thus, with the same definition of V as above, V nowsimply contains an extra term associated with the input u

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (154)

140 Advanced Stability Theory Chap. 4

V = x r P (Ax + Bu) = xrPBu - 1 xTQx (4.47)

Since y = Cx, we see that (4.47) defines a dissipative mapping between u and y,provided that the matrices B and C be related by

IThis result, known as the Kalman-Yakubovich (KY) lemma, shows the closeness of thepassivity and stability concepts, given compatible choices of inputs and outputs. Sincethe Lyapunov equation (4.46) can be satisfied for any arbitrary symmetric p.d. matrixQ, the KY lemma states that given any open-loop strictly stable linear system, one canconstruct an infinity of dissipative input-output maps simply by using compatiblechoices of inputs and outputs. In particular, given the system's physical inputs and theassociated matrix B, one can choose an infinity of outputs from which the linearsystem will look dissipative.

In the single-input case, and given our earlier discussion of frequency-domaincharacterizations of the passivity of linear systems, the KY lemma can be equivalentlystated as Lemma 4.4. Note that the controllability condition in that frequency-domainformulation simply ensures that the transfer function h(p) completely characterizes thelinear system defined by (A, b, c) (since P is symmetric positive definite and (A, b) iscontrollable, thus (A, cT) = (A, b^P) is observable). Also, as noted earlier, the KYlemma can be extended to PR systems, for which it can be shown that there exist apositive definite matrix P and a positive .semi-definite matrix Q such that (4.36a) and(4.36b) are verified. The main usefulness of this result is that it is applicable totransfer functions containing a pure integrator, which are common in adaptivecontroller design.

Example 4.20: The passivity of the adaptation law (4.40) of Example 4.19 can also be showndirectly by noticing that the integrator structure

9 = - e w(t)

implies that the mapping - e w{i) —» 9 is passive, and therefore that the mapping e —> - w(t) 8

is also passive (since 9 [ - e w(t)} = [ - w(t) 9] e ).

Furthermore, the passivity interpretation shows that the integrator in the above update law

can be replaced by any PR transfer function, while still guaranteeing that the tracking error e

tends to zero. Indeed, since the dissipation term #2 ' s simply zero using the original update law,

the KY lemma shows that, with the modified update law, there exists a symmetric positive

definite matrix P and a symmetric positive semi-definite matrix Q (which, in this simple first-

order case, are simply scalars P > 0 and Q > 0 ) such that

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (155)

Sect. 4.7 The Passivity Formalism 141

V = e2 + PQ2+\ Q[&(r)]2dr^ o

V = -2e2

The tracking convergence proof can then be completed as before using Barbalat's lemma.

Thus, the passivity interpretation can quickly suggest additional design flexibility. D

PASSIVITY INTERPRETATION OF SPR SYSTEMS

The passivity interpretation of SPR systems may allow one to quickly determinewhether a transfer function is SPR, using physical insights.

Consider for instance the transfer function

10/7h5(p) =

4p2 + 5p+

We can determine whether h5 is SPR using a procedure similar to that used for thefunction /z4 in Example 4.15. We can also simply notice that h5 can be interpreted asthe transfer function of a mass-spring-damper system

A'x + 5'x + x = 10M

y='x

with force as input and velocity as output. Thus h5 is dissipative, and thus SPR (sinceit is also strictly stable).

Finally, one can easily verify that

• If h(p) is SPR, so is l/h(p)

• If h\(p) and h2(p) are SPR, so is

h(p) = a, h{(p) + a2h2(p)

provided that ocj > 0 and a2 > 0

• If h\{p) and h2(p) are SPR, so is

h\(p)

\+hl{p)h2(p)

which is the overall transfer function of the negative feedback system havingh\(p) as the forward path transfer function and h2(p) as the feedback path

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (156)

142 Advanced Stability Theory

transfer function

Chap. 4

While these results can be derived directly, they are simply versions specific to thelinear single-input case of more general properties of passive systems. The first resultsimply reflects the fact that the input u and the output y play a symmetric role in thedefinition of passivity. The last two results represent the linear single-input frequency-domain interpretation of our earlier general discussion on the combination of passiveblocks. Actually, if either hx(p) or h2{p) is SPR, while the other is merely passive,then h(p) is SPR. This allows us to easily construct new SPR transfer functionssimply by taking any stable transfer function having a phase shift smaller than 90° atall frequencies, and putting it in feedback or in parallel configuration with any SPRtransfer function.

4.8 * Absolute Stability

The systems considered in this section have the interesting structure shown in Figure4.8. The forward path is a linear time-invariant system, and the feedback part is amemoryless nonlinearity, i.e., a nonlinear static mapping. The equations of suchsystems can be written as

x = A x - (4.48a)

(4.48b)

where (j> is some nonlinear function and G(p) = cT[pl- A] lb . Many systems ofpractical interest can be represented in this structure.

e

J *G(p)

y

Figure 4.8 : System structure in absolutestability problems

THE ISSUE OF ABSOLUTE STABILITY

The nonlinear system in Figure 4.8 has a special structure. If the feedback path simplycontains a constant gain, i.e., if <j)(v) = a y, then the stability of the whole system, alinear feedback system, can be simply determined by examining the eigenvalues of theclosed-loop system matrix A - ocbc^. However, the stability analysis of the wholesystem with an arbitrary nonlinear feedback function <j) is much more difficult.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (157)

Sect. 4.8 ' Absolute Stability 143

In analyzing this kind of system using Lyapunov's direct method, we usuallyrequire the nonlinearity to satisfy a so-called sector condition, whose definition isgiven below.

Definition 4.12 A continuous function 0 is said to belong to the sector [ky, k-^, ifthere exists two non-negative numbers k\ and £2 such that

=> kx^f<_k2 (4.49)

Geometrically, condition (4.49) implies that the nonlinear function always liesbetween the two straight lines k^y and k2y, as shown in Figure 4.9. Two properties areimplied by equation (4.49). First, it implies that (j)(0) = 0. Secondly, it implies thatyty(y) > 0, i.e, that the graph of ty(y) lies in the first and third quadrants. Note that inmany of later discussions, we will consider the special case of §(y) belonging to thesector [0, it], i.e, 3 k > 0, such that

0 < <j>(y) < ky (4.50)

Slope

Figure 4.9 : The sector condition (4.49)

Assume that the nonlinearity §(y) is a function belonging to the sector [yfcj, £2],and that the A matrix of the linear subsystem in the forward path is stable (i.e.,Hurwitz). What additional constraints are needed to guarantee the stability of thewhole system? In view of the fact that the nonlinearity in Figure 4.9 is bounded by thetwo straight lines, which correspond to constant gain feedback, it may be plausiblethat the stability of the nonlinear system should have some relation to the stability ofconstant gain feedback systems. In 1949, the Soviet scientist M.A. Aizerman madethe following conjecture: // the matrix [A — bc^k] is stable for all values of k in[&j, £2]> then the nonlinear system is globally asymptotically stable.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (158)

144 Advanced Stability Theory Chap. 4

POPOV'S CRITERION

Aizerman's is a very interesting conjecture. If it were true, it would allow us to deducethe stability of a nonlinear system by simply studying the stability of linear systems.However, several counter-examples show that this conjecture is false. AfterAizerman, many researchers continued to seek conditions that guarantee the stabilityof the nonlinear system in Figure 4.8. Popov's criterion imposes additional conditionson the linear subsystem, leading to a sufficient condition for asymptotic stabilityreminiscent of Nyquist's criterion (a necessary and sufficient condition) in linearsystem analysis.

A number of versions have been developed for Popov's criterion. The followingbasic version is fairly simple and useful.

Theorem 4.12 (Popov's Criterion) / / the system described by (4.48) satisfies theconditions

• the matrix A is Hurwitz (i.e., has all its eigenvalues strictly in the left half-plane) and the pair [A, b] is controllable

• the nonlinearity <|> belongs to the sector [0, k]

• there exists a strictly positive number a such that

V©>0 Re[(l +ja(o) G(yco)] + - > e (4.51)k

for an arbitrarily small e > 0, then the point 0 is globally asymptotically stable.

Inequality (4.51) is called Popov's inequality. The criterion can be provenconstructing a Lyapunov function candidate based on the KY lemma.

Let us note the main features of Popov's criterion:

• It only applies to autonomous systems.

• It is restricted to a single memoryless nonlinearity.

• The stability of the nonlinear system may be determined by examining thefrequency-response functions of a linear subsystem, without the need ofsearching for explicit Lyapunov functions.

• It only gives a sufficient condition.

The criterion is most easy to apply by using its graphical interpretation. Let

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (159)

Sect. 4.8 * Absolute Stability 145

Then expression (4.51) can be written

G1(co)-aa»G2(co) + - > e (4.52)K

Now let us construct an associated transfer function W(y'co), with the same real part asG(J co), but an imaginary part equal to co Im(G(y co)), i.e.,

W(/co) = x+jy= G^co) + ycoG2(co)

Then (4.52) implies that the nonlinear system is guaranteed to be globallyasymptotically stable if, in the complex plane having x and y as coordinates, the polarplot of W(j(o) is (uniformly) below the line x-ay + (l/k) = 0 (Figure 4.10). Thepolar plot of W is called a Popov plot. One easily sees the similarity of this criterionto the Nyquist criterion for linear systems. In the Nyquist criterion, the stability of alinear feedback system is determined by examining the position of the polar plot ofGQ'co) relative to the point (0, - 1), while in the Popov criterion, the stability of anonlinear feedback system is determined by checking the position of the associatedtransfer function W(j<$) with respect to a line.

Im

Figure 4.10 : The Popov plot

Example 4.21: Let us determine the stability of a nonlinear system of the form (4.48) where thelinear subsystem is defined by

p 2 + 7p+ 10

and the nonlinearity satisfies condition (4.50).

First, the linear subsystem is strictly stable, because its poles are - 2 and - 5 . It is also

controllable, because there is no pole-zero cancellation. Let us now check the Popov inequality.

The frequency response function O(yto) is

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (160)

146 Advanced Stability Theory Chap. 4

- « 2 + T

Therefore,

„ 4co2 + 3

o)4 + 29to2

„ -co(co2 +2 (B4 + 29(fl2

'0)7+ 10

SO

+ 100

H)+ 100

Substituting the above into (4.52) leads to

4 « 2 + 30+occo2(co2+ll)+(I -e)((04 + 29a>2+ 100)>0k

It is clear that this inequality can be satisfied by any strictly positive number a, and any strictly

positive number k, i.e., 0 < it < °° . Thus, the nonlinear system is globally asymptotically stable as

long as the nonlinearity belongs to the first and third quadrants. D

THE CIRCLE CRITERION

A more direct generalization of Nyquist's criterion to nonlinear systems is the circlecriterion, whose basic version can be stated as follows.

Theorem 4.13 (Circle Criterion) If the system (4.48) satisfies the conditions

• the matrix A has no eigenvalue on the ja> axis, and has p eigenvaluesstrictly in the right half-plane;

• the nonlinearity (j) belongs to the sector [/cj ,k2] >'

• one of the following is true

• 0 < kx < k2 , the Nyquist plot of G(j(a) does not enter the diskD(fej, kj) and encircles it p times counter-clockwise;

• 0 = &| < &2 , and the Nyquist plot ofG(j(a) stays in the half-plane

Rep>-l/k2 ;

• ky <0<k2, and the Nyquist plot ofG(j(n) stays in the interior ofthediskT>(kx,k2) ;

• ki<k2<0 , the Nyquist plot of - G(j(i>) does not enter the diskD(—k\,-k2) and encircles it p times counter-clockwise;

then the equilibrium point 0 of the system is globally asymptotically stable.

j

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (161)

Sect. 4.9 * Establishing Boundedness of Signals 147

Im G(jo>)

ReG(jm)

Figure 4.11 : The circle criterion

Thus we see that, essentially, the critical point - l/k in Nyquist's criterion isreplaced in the circle criterion by the circle of Figure 4.11 (which tends towards thepoint - l/£j as k2 tends to £j, i.e., as the conic sector gets thinner). Of course, thecircle criterion states sufficient but not necessary conditions.

The circle criterion can be extended to non-autonomous systems.

4.9 * Establishing Boundedness of Signals

In the stability analysis or convergence analysis of nonlinear systems, a frequentlyencountered problem is that of establishing the boundedness of certain signals. Forinstance, in order to use Barbalat's lemma, one has to show the uniform continuity of/, which can be most conveniently shown by proving the boundedness of / . Similarly,in studying the effects of disturbances, it is also desirable to prove the boundedness ofsystem signals in the presence of disturbances. In this section, we provide two usefulresults for such purposes.

THE BELLMAN-GRONWALL LEMMA

In system analysis, one can often manipulate the signal relations into an integralinequality of the form

y(t) < \' a{x)y(x)dT + b{t) (4.53)

where y(t), the variable of concern, appears on both sides of the inequality. Theproblem is to gain an explicit bound on the magnitude of y from the above inequality.The Bellman-Gronwall lemma can be used for this purpose.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (162)

148 Advanced Stability Theory Chap. 4

Lemma 4.6 (Bellman-Gronwall) Let a variable y(t) satisfy (4.53), with aft), bft)being known real functions. Then

y(t) < [' a(T)b(%) exp[\' a(r)dr]dx + b{t) (4.54)O JT

Ifb(t) is differentiable, then

yit) < fe(O) exp[f'a(t)dx]+ \'b{x) exp[['a(r)dr)dx (4.55)JO JO J T

In particular, if bft) is a constant, we simply have

y(t)<b(O)exptf'a(x)dx] (4.56)

Proof: The proof is based on defining a new variable and transforming the integral inequality into

a differential equation, which can be easily solved. Let

v(()=f a(x)y(x)dx (4.57)o

Then differentiation of v and use of (4.53) leads to

v = a(t)y(t) < a(t) v(r) + a(f) b{i)

Let

i(/) = a(t) y(t) - a(t) v(i)-a(t) b(t)

which is obviously a non-positive function. Then v(0 satisfies

v(0 - a{t) v(t) = a(t)b(t) + 5(0

Solving this equation with initial condition v(0) = 0, yields

v(0= \'exp[\'a(r)dr] [a(x)b(x) + s(x)]dx (4.58)

Since 5() is a non-positive function,

r' r 'v (0< | exp[J a(r)dr] a(x) b(x)dx

This, together with the definition of v and the inequality (4.53), leads to

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (163)

Sect. 4.9 * Establishing Boundedness of Signals 149

y(t) < f'exp[['<*(/•)dr] a(x)b(x)dx + b(t)O X

If /)(() is differentiable, we obtain, by partial integration

f exp[f a(r)dr]a(x)b(x)dx = - fc(t)exp[f a(r)dr] | IZ' + f'b(z)exp[('a(r)dr)di O

Jn Jt JT Ja JX

TOTAL STABILITY

Consider the nonlinear system3x = d(t) (4.59)

which represents a non-linear mass-spring-damper system with a disturbance d(t)(which may be due to unmodeled Coulomb friction, motor ripple, parametervariations, etc). Is x bounded when the disturbance is bounded? This is the mainquestion addressed by the so-called total stability theory (or stability under persistentdisturbances).

In total stability, the systems concerned are described in the form

x = f(x, 0 + g(x, t) (4.60)

where g(x,/) is a perturbation term. Our objective is to derive a boundedness conditionfor the perturbed equation (4.60) from the stability properties of the associatedunperturbed system

x = f(x, t) (4.61)

We assume that x = 0 is an equilibrium point for the unperturbed dynamics (4.61), i.e.,f(0, t) = 0. But the origin is not necessarily an equilibrium point for the perturbeddynamics (4.60). The concept of total stability characterizes the ability of a system towithstand small persistent disturbances, and is defined as follows:

Definition 4.13 The equilibrium point x = 0 for the unperturbed system (4.61) issaid to be totally stable if for every e > 0, two numbers 8y and 82 exist such that||x(/o)||<8j and ||g(x,r)|| < 52 imply that every solution x(t) of the perturbed system

(4.60) satisfies the condition \\x(t)\\ < £.

The above definition means that an equilibrium point is totally stable if the stateof the perturbed system can be kept arbitrarily close to zero by restricting the initialstate and the perturbation to be sufficiently small. Note that total stability is simply alocal version (with small input) of BIBO (bounded input bounded output) stability. It

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (164)

150 Advanced Stability Theory Chap. 4

is also useful to remark that if the unperturbed system is linear, then total stability isguaranteed by the asymptotic stability of the unperturbed system.

The following theorem is very useful to assert the total stability of a nonlinearsystem.

Theorem 4.14 If the equilibrium point of (4.61) is uniformly asymptotically stable,then it is totally stable.

This theorem can be proven by using the converse Lyapunov theorem 4.8. It meansthat uniformly asymptotically stable systems can withstand small disturbances.Because uniformly asymptotic stability can be asserted by the Lyapunov theorem 4.1,total stability of a system may be similarly established by theorem 4.1. Note thatasymptotic stability is not sufficient to guarantee the total stability of a nonlinearsystem as can be verified by counter-examples. We also point out that exponentiallystable systems are always totally stable because exponential stability implies uniformasymptotic stability.

Example 4.22: Consider again the system (4.59). Let us analyze the stability of the unperturbed

system

x + 2 i 3 + 3x = 0

first. Using the scalar function

and the invariant set theorem, one easily shows that the equilibrium point is globally

asymptotically stable. Because the system is autonomous, the stability is also uniform. Thus, the

above theorem shows that the system can withstand small disturbances d(t). D

Total stability guarantees boundedness to only small-disturbance, and requiresonly local uniform asymptotic stability of the equilibrium point. One might wonderwhether the global uniform asymptotic stability can guarantee the boundedness of thestate in the presence of large (though still bounded) perturbations. The followingcounter-example demonstrates that this is not true.

Example 4.23: The nonlinear equation

= w(t) (4.62)

can be regarded as representing mass-spring-damper system containing nonlinear damping fix)

and excitation force w(t), where / is a first and third quadrant continuous nonlinear function such

that

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (165)

Sect. 4.10

\f(y) I < 1 _ oo < _y < oo

as illustrated in Figure 4.12.

' Existence and Unicity of Solutions 151

fly)

Figure 4.12 : A nonlinear damping

function

The system is totally stable, because the equilibrium point can be shown to be globally uniformly

asymptotically stable using the Lyapunov function V = (1/2)(x2 + x2). Is the output bounded for

bounded input?

Let us consider the response of the system to the excitation force w(t) = A sin /, A> S/n. By

writing (4.62) as

x + x = Asint-f{x)

and solving this linear equation with (A sin t -f(x)) as input, we obtain

A c* . A rl

x(t) = —(sint-tcost)-\ sin(t-x)f<Jc)d% > — (s'mt-tcost) - I \s'm(t-T)\dx2 J0 2 J0

The integral on the right-hand side can be shown to be smaller than (2/;i)(l + e)t for any l> tg,

and e > 0 and some to . Attn = (2« + l)7t,

x(tn)>

Therefore, if we'take A > 8/71 and e = 1/2, x(tn) —> 0°. •4.10 * Existence and Unicity of Solutions

This section discusses the mathematically important question of existence and unicityof solutions of nonlinear differential equations. We first describe a simple and quitegeneral sufficient condition for a nonlinear differential equation to admit a solution,and then simple but conservative conditions for this solution to be unique.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (166)

152 Advanced Stability Theory Chap. 4

Theorem 4.15 (Cauchy existence theorem) Consider the differential equationx = f(x, t), with initial condition \(t0) = x0. If the function f is continuous in the closedregion

\t-to\ < T, \\x-xo\\ < R

where T and R are strictly positive constants, then the equation has at least onesolution x(t) which satisfies the initial condition and is continuous over a finite timeperiod [to,t{] (where tx > to).

The above theorem indicates that the continuity of f is sufficient for the localexistence of solutions. However, it does not guarantee the uniqueness of the solution.

Example 4.24: An equation with multiple solutions

Consider the equation

with initial condition y(Q) = 0. Two of its solutions are y(t) = 0 and y = r3. [U

The following theorem gives a sufficient condition for the unique existence of asolution.

Theorem 4.16 If the function f(x, t) is continuous in t, and if there exists a strictlypositive constant L such that

||f(x2,r)-f(x,,r)|| < L \\x2-xx\\ (4.63)

for all Xy and X2 in a finite neighborhood of the origin and all t in the interval\t0, to + T\ (with T being a strictly positive constant), then x — f(x, t) has a uniquesolution x(t)for sufficiently small initial states and in a sufficiently short time interval.

Condition (4.63) is called a Lipschitz condition and L is known as a Lipschitzconstant. If (4.63) is verified, then f is said to be locally Lipschitz in x. If (4.63) isverified for any time t, then f is said to be locally Lipschitz in x uniformly with respectto t. Note that the satisfaction of a Lipschitz condition implies (locally) the continuityof f in terms of x, as can be easily proven from the definition of continuity.Conversely, if locally f has a continuous and bounded Jacobian with respect to x, thenf is locally Lipschitz. When (4.63) is satisfied for any x( and x2 in the state space, f issaid to be globally Lipschitz. The above theorem can then be extended to guaranteeunique existence of a solution in a global sense (i.e., for any initial condition and anytime period).

While the condition for existence of solutions, as stated by Cauchy's theorem, is

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (167)

Sect. 4.12 Notes and References 153

rather benign, the sufficient condition for unicity is quite strong, and, actually, overlyconservative. Most results on nonlinear dynamics simply assume that f is smoothenough to guarantee existence and unicity of the solutions. Note that this is always thecase of good physical system models (at least in classical physics).

Actually, precise mathematical results exist about the relation between theexistence of a Lyapunov function for a given system and the existence and unicity ofsolutions (see, e.g., [Yoshizawa, 1966, 1975]). From a practical point of view, theseresults essentially mean that the existence of a Lyapunov function to describe a systemwill guarantee the system's "good behavior" under some mild smoothnessassumptions on the dynamics.

4.11 Summary

Some advanced topics in nonlinear control theory are presented in this chapter.Lyapunov theory for non-autonomous systems is discussed first. Its results are quitesimilar to those for autonomous systems, although more involved conditions arerequired. A major difference is that the powerful invariant-set theorem does not applyto non-autonomous system, although Barbalat's lemma can often be a simple andeffective substitute. A number of instability theorems are also presented. Suchtheorems are useful for non-autonomous systems, or for autonomous systems whoselinearizations are only marginally stable. Theorems on the existence of Lyapunovfunctions may be of use in constructing Lyapunov functions for systems part of whichis known to have certain stability properties. The passivity formalism is alsointroduced, as a notationally convenient and physically motivated interpretation ofLyapunov or Lyapunov-like analysis. The chapter also includes some results forestablishing the boundedness of signals in nonlinear systems.

4.12 Notes and References

A comprehensive yet readable book on Lyapunov analysis of non-autonomous systems is [Hahn,

1967], on which most of the stability definitions in this chapter are based. The definitions and

results concerning positive definite and decrescent functions are based on [Hahn, 1967; Vidyasagar,

1978]. The statement and proof of Theorem 4.1 are adapted from [Kalman and Bertram, I960].

Example 4.5 is adapted from [Vidyasagar, 1978], Example 4.6 from [Massera, 1949], and Example

4.7 from [Rouche, el a/., 1977]. Figure 4.1 is adapted from A.S.M.E. Journal of Basic Engineering,

1960. In section 4.2.2, the result on perturbed linear systems is from [Vidyasagar, 1978], while the

result on sufficient smoothness conditions on the A(;) matrix is from [Middleton, 1988]. Sections

4.2.3 and 4.3 are largely adapted from [Vidyasagar, 1978], where proofs of the main results can be

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (168)

154 Advanced Stability Theory Chap. 4

found. The statement of Theorem 4.9 follows that in [Bodson, 1986]. Lemma 4.2 and its proof are

from [Popov, 1973]. An extensive study of absolute stability problems from a frequency-domain

perspective is contained in [Narendra and Taylor, 1973], from which the definitions and theorems on

positive real functions are adapted. A more recent description of positive real functions and their

applications in adaptive control can be found in [Narendra and Annasswamy, 1989]. The Bellman-

Gronwall lemma and its proof are adapted from [Hsu and Meyer, 1968]. The definition and theorem

on total stability are based on [Hahn, 1965]. Example 4.23 is adapted from [Desoer, et al., 1965].

Passivity theory (see [Popov, 1973; Desoer and Vidyasagar, 1975]) is presented in a slightly

unconventional form. Passivity interpretations of adaptive control laws are discussed in [Landau,

1979]. The reader is referred to [Vidyasagar, 1978] for a detailed discussion of absolute stability.

The circle criterion and its extensions to non-autonomous systems were derived by [Narendra and

Goldwyn, 1964; Sandberg, 1964; Tsypkin, 1964; Zames, 1966],

Other important robustness analysis tools include singular perturbations (see, e.g.,

[Kokotovic, et al., 1986]) and averaging (see, e.g, [Hale, 1980; Meerkov, 1980]).

Relations between the existence of Lyapunov functions and the existence and unicity of

solutions of nonlinear differential equations are discussed in [Yoshizawa, 1966, 1975].

4.13 Exercises

4.1 Show that, for a non-autonomous system, a system trajectory is generally not an invariant set.

4.2 Analyze the stability of the dynamics (corresponding to a mass sinking in a viscous liquid)

v + 2a|v|v + bv = c a>0,b>0

4.3 Show that a function V(x, t) is radially unbounded if, and only if, there exists a class-K

function <j> such that

V(x,0><KIMI)

where the function (j) satisfies

Urn <KIM|) = oo

4.4 The performance of underwater vehicles control systems is often constrained by the

"unmodeled" dynamics of the thrusters. Assume that one decides to explicitly account for thruster

dynamics, based on the model

d) = - a , co|a)| + a 2 x a ! > 0 , a - , > 0

u = ba> |oo| b>0

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (169)

Sect. 4.13 Exercises 155

where x is the torque input to the propeller, co is the propeller's angular velocity, and » is the actual

thrust generated.

Show that, for a constant torque input x0 , the steady-state thrust is proportional to xg (which is

consistent witht the fact that thruster dynamics is often treated as "unmodeled").

Assuming that the coefficients a,- and b in the above model are known with good accuracy,

design and discuss the use of a simple "open-loop" observer for u (given an arbitrary time-varying

torque input x) in the absence of measurements of (0. (Adapted from [Yoerger and Slotine, 1990].)

4.5 Discuss the similarity of the results of section 4.2.2 with Krasovskii's theorem of section 3.5.2.

4.6 Use the first instability theorem to show the instability of the vertical-up position of a

pendulum.

4.7 Show explicitly why the linear time-varying system defined by (4.18) does not satisfy the

sufficient condition (4.19).

4.8 Condition (4.19) on the eigenvalues of A(t) + AT(t) is only, of course, a sufficient condition.

For instance, show that the linear time-varying system associated with the matrix

-1 e"2

0 -1

is globally asymptotically stable.

4.9 Determine whether the following systems have a stable equilibrium. Indicate whether the

stability is asymptotic, and whether it is global.

(a)- 10 e3'

0 - 2

(b)x\ -1 2sinr

0 - ( / + ! ) X2

x\=

' - 1 e2''

0 - 2

x\

X2

(c)

4.10 If a differeniiable function/is lower bounded and decreasing ( / < 0), then it converges to a

limit. However,/does not necessarily converge to zero. Derive a counter-example. (Hint: You may

L

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (170)

156 Advanced Stability Theory Chap. 4

use for —/ a function that peaks periodically, but whose integral is finite.)

4.11 (a) Show that if a function/is bounded and uniformly continuous, and there exists a positive

definite function F(f, t) such that

J F(f(t), t) dt < o o

1then f(t) tends to zero as f —> °°.

(b) For a given autonomous nonlinear system, consider a Lyapunov function Vina ball B^ ,

and let $ be a scalar, differentiable, strictly monotonously increasing function of its scalar argument.

Show that [<|>(V)-<|>(0)] is also a Lyapunov function for the system (distinguish the cases of

stability and of asymptotic stability). Suggest extensions to non-autonomous systems.

4.12 Consider a scalar, lower bounded, and twice continuously differentiable function V(t) such

that

V / > 0 , V(0<0

Show that, for any / > 0 ,

V(0 = 0 => V(t) = 0

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (171)

Chapter 5Describing Function Analysis

The frequency response method is a powerful tool for the analysis and design of linearcontrol systems. It is based on describing a linear system by a complex-valuedfunction, the frequency response, instead of a differential equation. The power of themethod comes from a number of sources. First, graphical representations can be usedto facilitate analysis and design. Second, physical insights can be used, because thefrequency response functions have clear physical meanings. Finally, the method'scomplexity only increases mildly with system order. Frequency domain analysis,however, cannot be directly applied to nonlinear systems because frequency responsefunctions cannot be defined for nonlinear systems.

Yet, for some nonlinear systems, an extended version of the frequency responsemethod, called the describing function method, can be used to approximately analyzeand predict nonlinear behavior. Even though it is only an approximation method, thedesirable properties it inherits from the frequency response method, and the shortageof other systematic tools for nonlinear system analysis, make it an indispensablecomponent of the bag of tools of practicing control engineers. The main use ofdescribing function method is for the prediction of limit cycles in nonlinear systems,although the method has a number of other applications such as predictingsubharmonics, jump phenomena, and the response of nonlinear systems to sinusoidalinputs.

This chapter presents an introduction to the describing function analysis ofnonlinear systems. The basic ideas in the describing function method are presented in

157

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (172)

158 Describing Function Analysis Chap. 5

section 5.1. Section 5.2 discusses typical "hard nonlinearities" in control engineering,since describing functions are particularly useful for studying control systemscontaining such nonlinearities. Section 5.3 evaluates the describing functions forthese hard nonlinearities. Section 5.4 is devoted to the description of how to use thedescribing function method for the prediction of limit cycles.

5.1 Describing Function Fundamentals

In this section, we start by presenting describing function analysis using a simpleexample, adapted from [Hsu and Meyer, 1968]. We then provide the formal definitionof describing functions and some techniques for evaluating the describing functions ofnonlinear elements.

5.1.1 An Example of Describing Function Analysis

The interesting and classical Van der Pol equation

x + a(x2-\)x + x = 0 (5.1)

(where a is a positive constant) has been treated by phase-plane analysis andLyapunov analysis in the previous chapters. Let us now study it using a differenttechnique, which shall lead us to the concept of a describing function. Specifically, letus determine whether there exists a limit cycle in this system and, if so, calculate theamplitude and frequency of the limit cycle (pretending that we have not seen the phaseportrait of the Van der Pol equation in Chapter 2). To this effect, we first assume theexistence of a limit cycle with undetermined amplitude and frequency, and thendetermine whether the system equation can indeed sustain such a solution. This isquite similar to the assumed-variable method in differential equation theory, where wefirst assume a solution of certain form, substitute it into the differential equation, andthen attempt to determine the coefficients in the solution.

Before carrying out this procedure, let us represent the system dynamics in ablock diagram form, as shown in Figure 5.1. It is seen that the feedback system in 5.1contains a linear block and a nonlinear block, where the linear block, althoughunstable, has low-pass properties.

Now let us assume that there is a limit cycle in the system and the oscillationsignal x is in the form of

x(t) =/\sin(coO

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (173)

Sect. 5.1 Describing Function Fundamentals

Nonlinear Element ( - x x2) Linear Element

159

Figure 5.1 : Feedback interpretation of the Van der Pol oscillator

with A being the limit cycle amplitude and co being the frequency. Thus,

x(t) =Acocos(a>0

Therefore, the output of the nonlinear block is

w = -x2x = -A2sin2((Of) Atocos(toz)

( l - cos(2coO ) cos(tor) = - d _ ^ (cos(a>0 - cos(3a>0 )

It is seen that w contains a third harmonic term. Since the linear block has low-passproperties, we can reasonably assume that this third harmonic term is sufficientlyattenuated by the linear block and its effect is not present in the signal flow after thelinear block. This means that we can approximate w by

A3 A2 A

w = - — cocosccw = [-Asin(cor)]4 4 d!

so that the nonlinear block in Figure 5.1 can be approximated by the equivalent"quasi-linear" block in Figure 5.2. The "transfer function" of the quasi-linear blockdepends on the signal amplitude A, unlike a linear system transfer function (which isindependent of the input magnitude).

In the frequency domain, this corresponds to

w = N(A, co) ( - x ) (5.2)

where

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (174)

160 Describing Function Analysis Chap. 5

,•=0

QUASI-LINEARAPPROXIMATION

1

A2

— P4

1 |

w a

P2- ap +1

X

Figure 5.2 : Quasi-linear approximation of the Van der Pol oscillator

N(A, co) = ^ ( j

That is, the nonlinear block can be approximated by the frequency response functionN(A, co). Since the system is assumed to contain a sinusoidal oscillation, we have

x = A sin(cor) = G(j(n) w = G(jco) N(A, co) ( - x )

where G(y'co) is the linear component transfer function. This implies that

+ A2(jco) a = 04 C/co)2-a<jco) + l

Solving this equation, we obtain

A = 2 co= 1

Note that in terms of the Laplace variable p, the closed-loop characteristic equation ofthis system is

1 +- •P a4 p2 - ap +

whose eigenvalues are

= 0

X,,2 = - ia(A2-4) ± |±a2(A2-4)2 -

(5.3)

(5.4)

Corresponding to A = 2, we obtain the eigenvalues A,j2 = ±7- This indicates theexistence of a limit cycle of amplitude 2 and frequency 1. It is interesting to noteneither the amplitude nor the frequency obtained above depends on the parameter a in

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (175)

Sect. 5.1

Equation 5.1.

Describing Function Fundamentals 161

In the phase plane, the above approximate analysis suggests that the limit cycleis a circle of radius 2, regardless of the value of a. To verify the plausibility of thisresult, the real limit cycles corresponding to the different values of a are plotted(Figure 5.3). It is seen that the above approximation is reasonable for small value ofa, but that the inaccuracy grows as a increases. This is understandable because as agrows the nonlinearity becomes more significant and the quasi-linear approximationbecomes less accurate.

limit cycle

Figure 5.3 : Real limit cycles on the phase plane

The stability of the limit cycle can also be studied using the above analysis. Letus assume that the limit cycle's amplitude A is increased to a value larger than 2.Then, equation (5.4) shows that the closed-loop poles now have a negative real part.This indicates that the system becomes exponentially stable and thus the signalmagnitude will decrease. Similar conclusions are obtained assuming that the limitcycle's amplitude A is decreased to a value less than 2. Thus, we conclude that thelimit cycle is stable with an amplitude of 2.

Note that, in the above approximate analysis, the critical step is to replace thenonlinear block by the quasi-linear block which has the frequency response function(A2/4) (j(0). Afterwards, the amplitude and frequency of the limit cycle can bedetermined from 1 + G(ja>) N(A, co) = 0. The function N(A, co) is called the describingfunction of the nonlinear element. The above approximate analysis can be extended topredict limit cycles in other nonlinear systems which can be represented into the blockdiagram similar to Figure 5.1, as we shall do in section 5.4.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (176)

162 Describing Function Analysis

5.1.2 Applications Domain

Chap. 5

Before moving on to the formal treatment of the describing function method, let usbriefly discuss what kind of nonlinear systems it applies to, and what kind ofinformation it can provide about nonlinear system behavior.

THE SYSTEMS

Simply speaking, any system which can be transformed into the configuration inFigure 5.4 can be studied using describing functions. There are at least two importantclasses of systems in this category.

Nonlinear Element Linear Element

r(0 = 0w=f(x)

wft)G(p)

y(0

Figure 5.4 : A nonlinear system

The first important class consists of "almost" linear systems. By "almost" linearsystems, we refer to systems which contain hard nonlinearities in the control loop butare otherwise linear. Such systems arise when a control system is designed usinglinear control but its implementation involves hard nonlinearities, such as motorsaturation, actuator or sensor dead-zones, Coulomb friction, or hysteresis in the plant.An example is shown in Figure 5.5, which involves hard nonlinearities in the actuator.

Example 5.1: A system containing only one nonlinearity

Consider the control system shown in Figure 5.5. The plant is linear and the controller is also

linear. However, the actuator involves a hard nonlinearity. This system can be rearranged into the

form of Figure 5.4 by regarding G

nonlinearity as the nonlinear element.

as the linear component G, and the actuator

D

"Almost" linear systems involving sensor or plant nonlinearities can besimilarly rearranged into the form of Figure 5.4.

The second class of systems consists of genuinely nonlinear systems whosedynamic equations can actually be rearranged into the form of Figure 5.4. We saw anexample of such systems in the previous section.

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (177)

Sect. 5.1

G/P)

Describing Function Fundamentals 163

w(t) u(t)Gp(p)

y(t)

Figure 5.5 : A control system with hard nonlinearity

APPLICATIONS OF DESCRIBING FUNCTIONS

For systems such as the one in Figure 5.5, limit cycles can often occur due to thenonlinearity. However, linear control cannot predict such problems. Describingfunctions, on the other hand, can be conveniently used to discover the existence oflimit cycles and determine their stability, regardless of whether the nonlinearity is"hard" or "soft." The applicability to limit cycle analysis is due to the fact that theform of the signals in a limit-cycling system is usually approximately sinusoidal. Thiscan be conveniently explained on the system in Figure 5.4. Indeed, asssume that thelinear element in Figure 5.4 has low-pass properties (which is the case of mostphysical systems). If there is a limit cycle in the system, then the system signals mustall be periodic. Since, as a periodic signal, the input to the linear element in Figure 5.4can be expanded as the sum of many harmonics, and since the linear element, becauseof its low-pass property, filters out higher frequency signals, the output y(t) must becomposed mostly of the lowest harmonics. Therefore, it is appropriate to assume thatthe signals in the whole system are basically sinusoidal in form, thus allowing thetechnique in subsection 5.1.1 to be applied.

Prediction of limit cycles is very important, because limit cycles can occurfrequently in physical nonlinear system. Sometimes, a limit cycle can be desirable.This is the case of limit cycles in the electronic oscillators used in laboratories.Another example is the so-called dither technique which can be used to minimize thenegative effects of Coulomb friction in mechanical systems. In most control systems,however, limit cycles are undesirable. This may be due to a number of reasons:

1. limit cycle, as a way of instability, tends to cause poor control accuracy

2. the constant oscillation associated with the limit cycles can causeincreasing wear or even mechanical failure of the control systemhardware

3. limit cycling may also cause other undesirable effects, such as passenger

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (178)

164 Describing Function Analysis Chap. 5

discomfort in an aircraft under autopilot

In general, although a precise knowledge of the waveform of a limit cycle is usuallynot mandatory, the knowledge of the limit cycle's existence, as well as that of itsapproximate amplitude and frequency, is critical. The describing function method canbe used for this purpose. It can also guide the design of compensators so as to avoidlimit cycles.

5.1.3 Basic Assumptions

Consider a nonlinear system in the general form of Figure 5.4. In order to develop thebasic version of the describing function method, the system has to satisfy thefollowing four conditions:

1. there is only a single nonlinear component

2. the nonlinear component is time-invariant

3. corresponding to a sinusoidal input x= sin(coO , only the fundamentalcomponent Wj(f) in the output w(t) has to be considered

4. the nonlinearity is odd

The first assumption implies that if there are two or more nonlinear componentsin a system, one either has to lump them together as a single nonlinearity (as can bedone with two nonlinearities in parallel), or retain only the primary nonlinearity andneglect the others.

The second assumption implies that we consider only autonomous nonlinearsystems. It is satisfied by many nonlinearities in practice, such as saturation inamplifiers, backlash in gears, Coulomb friction between surfaces, and hysteresis inrelays. The reason for this assumption is that the Nyquist criterion, on which thedescribing function method is largely based, applies only to linear time-invariantsystems.

The third assumption is the fundamental assumption of the describing functionmethod. It represents an approximation, because the output of a nonlinear elementcorresponding to a sinusoidal input usually contains higher harmonics besides thefundamental. This assumption implies that the higher-frequency harmonics can all beneglected in the analysis, as compared with the fundamental component. For thisassumption to be valid, it is important for the linear element following the nonlinearityto have low-pass properties, i.e.,

|G(jco)| » | G(;«co) | for n = 2, 3,... (5.5)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (179)

Sect. 5.1 Describing Function Fundamentals 165

This implies that higher harmonics in the output will be filtered out significantly.Thus, the third assumption is often referred to as the filtering hypothesis.

The fourth assumption means that the plot of the nonlinearity relation f(x)between the input and output of the nonlinear element is symmetric about the origin.This assumption is introduced for simplicity, i.e., so that the static term in the Fourierexpansion of the output can be neglected. Note that the common nonlinearitiesdiscussed before all satisfy this assumption.

The relaxation of the above assumptions has been widely studied in literature,leading to describing function approaches for general situations, such as multiplenonlinearities, time-varying nonlinearities, or multiple-sinusoids. However, thesemethods based on relaxed conditions are usually much more complicated than thebasic version, which corresponds to the above four assumptions. In this chapter, weshall mostly concentrate on the basic version.

5.1.4 Basic Definitions

Let us now discuss how to represent a nonlinear component by a describing function.Let us consider a sinusoidal input to the nonlinear element, of amplitude A andfrequency co, i.e., x(t) = Asin(au), as shown in Figure 5.6. The output of the nonlinearcomponent w{i) is often a periodic, though generally non-sinusoidal, function. Notethat this is always the case if the nonlinearity fix) is single-valued, because the outputis/[Asin(oo(/+2n:/GO))] =/[Asin(cor)]. Using Fourier series, the periodic function w(t)can be expanded as

oo

w(0 = —+ ~^[ancos(n(i)t) + bnsin(n(i)t)] (5.6)

where the Fourier coefficients a('s and bfs are generally functions of A and (0,determined by

ao = -\* w(t)d((M) (5.7a)

a = I f * w(t)cos (no)t)d(at) (5.7b)KJ-K

bn = -\K w(t)sm(n(nt)d((tit) (5.7c)ft - i t

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (180)

166 Describing Function Analysis Chap. 5

A sin(tt) t)N.L.

w(t) A sin((0 t)N(A, w)

M sin

Figure 5.6 : A nonlinear element and its describing function representation

Due to the fourth assumption above, one has a0 = 0. Furthermore, the thirdassumption implies that we only need to consider the fundamental component wx(t),namely

w(t) ~ wj(f) = aj cos(cof) + bx sin(a>0 = Msin(co? + (()) (5.8)

where

M{A,co) = J a ^ + 6,2 and <j>(A,co) =

Expression (5.8) indicates that the fundamental component corresponding to asinusoidal input is a sinusoid at the same frequency. In complex representation, thissinusoid can be written as wj = Mei(m + ^ = (b{ +ja\) eJmt.

Similarly to the concept of frequency response function, which is the frequency-domain ratio of the sinusoidal input and the sinusoidal output of a system, we definethe describing function of the nonlinear element to be the complex ratio of thefundamental component of the nonlinear element by the input sinusoid, i.e.,

(pi \x i

/V(A,co)=^ifl^ ^e^=Ub{+iax) (5.9)

With a describing function representing the nonlinear component, the nonlinearelement, in the presence of sinusoidal input, can be treated as if it were a linearelement with a frequency response function /V(A,co), as shown in Figure 5.6. Theconcept of a describing function can thus be regarded as an extension of the notion offrequency response. For a linear dynamic system with frequency response functionH(jw), the describing function is independent of the input gain, as can be easilyshown. However, the describing function of a nonlinear element differs from thefrequency response function of a linear element in that it depends on the inputamplitude A. Therefore, representing the nonlinear element as in Figure 5.6 is alsocalled quasi-linearization.

Generally, the describing function depends on the frequency and amplitude ofthe input signal. There are, however, a number of special cases. When thenonlinearity is single-valued, the describing function iV(A,co) is real and independentof the input frequency co. The realness of N is due to the fact that ax = 0, which is true

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (181)

Sect. 5.1 Describing Function Fundamentals 167

because f[A sin(a>0] cos(cof), the integrand in the expression (5.7b) for <Z[, is an oddfunction of cor, and the domain of integration is the symmetric interval [-71, TC]. Thefrequency-independent nature is due to the fact that the integration of the single-valued function f[A sin(cor)] sin(coO in expression (5.7c) is done for the variable cof,which implies that co does not explicitly appear in the integration.

Although we have implicitly assumed the nonlinear element to be a scalarnonlinear function, the definition of the describing function also applies to the casewhen the nonlinear element contains dynamics {i.e., is described by differentialequations instead of a function). The derivation of describing functions for suchnonlinear elements is usually more complicated and may require experimentalevaluation.

5.1.5 Computing Describing Functions

A number of methods are available to determine the describing functions of nonlinearelements in control systems, based on definition (5.9). We now briefly describe threesuch methods: analytical calculation, experimental determination, and numericalintegration. Convenience and cost in each particular application determine whichmethod should be used. One thing to remember is that precision is not critical inevaluating describing functions of nonlinear elements, because the describing functionmethod is itself an approximate method.

ANALYTICAL CALCULATION

When the nonlinear characteristics w =/(x) (where x is the input and w the output) ofthe nonlinear element are described by an explicit function and the integration in (5.7)can be easily carried out, then analytical evaluation of the describing function basedon (5.7) is desirable. The explicit function/(x) of the nonlinear element may be anidealized representation of simple nonlinearities such as saturation and dead-zone, or itmay be the curve-fit of an input-output relationship for the element. However, fornonlinear elements which evade convenient analytical expressions or containdynamics, the analytical technique is difficult.

NUMERICAL INTEGRATION

For nonlinearities whose input-output relationship w =f{x) is given by graphs ortables, it is convenient to use numerical integration to evaluate the describingfunctions. The idea is, of course, to approximate integrals in (5.7) by discrete sumsover small intervals. Various numerical integration schemes can be applied for thispurpose. It is obviously important that the numerical integration be easily

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (182)

168 Describing Function Analysis Chap. 5

implementable by computer programs. The result is a plot representing the describingfunction, which can be used to predict limit cycles based on the method to bedeveloped in section 5.4.

EXPERIMENTAL EVALUATION

The experimental method is particularly suitable for complex nonlinearities anddynamic nonlinearities. When a system nonlinearity can be isolated and excited withsinusoidal inputs of known amplitude and frequency, experimental determination ofthe describing function can be obtained by using a harmonic analyzer on the output ofthe nonlinear element. This is quite similar to the experimental determination offrequency response functions for linear elements. The difference here is that not onlythe frequencies, but also the amplitudes of the input sinusoidal should be varied. Theresults of the experiments are a set of curves on complex planes representing thedescribing function N(A, co), instead of analytical expressions. Specializedinstruments are available which automatically compute the describing functions ofnonlinear elements based on the measurement of nonlinear element response toharmonic excitation.

Let us illustrate on a simple nonlinearity how to evaluate describing functionsusing the analytical technique.

Example 5.2: Describing function of a hardening spring

The characteristics of a hardening spring are given by

w = x + x3/2

with x being the input and w being the output. Given an input x(t) = A sin(co(), the output

w(t) — A sin(cof) + A3 sin3(cof)/2 can be expanded as a Fourier series, with the fundamental being

W|(;) = tf| cos cor + foj sin co/

Because w(t) is an odd function, one has a^ = 0, according to (5.7). The coefficient ftj is

b,=-\ [Asin((Ot) + A3 sin3(mt)/2] sin((ot) d((Ot) = A + - A 3

Therefore, the fundamental is

w, =(A + -A3)sin(cor)8

and the describing function of this nonlinear component is

A

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (183)

Sect. 5.2 Common Nonlinearities In Control Systems 169

= 1 +-A2

8

Note that due to the odd nature of this nonlinearity, the describing function is real, being a

function only of the amplitude of the sinusoidal input. •

5.2 Common Nonlinearities In Control Systems

In this section, we take a closer look at the nonlinearities found in control systems.Consider the typical system block shown in Figure 5.7. It is composed of four parts: aplant to be controlled, sensors for measurement, actuators for control action, and acontrol law, usually implemented on a computer. Nonlinearities may occur in any partof the system, and thus make it a nonlinear control system.

y(t)controller actuators plant

Figure 5.7 : Block diagram of control systems

CONTINUOUS AND DISCONTINUOUS NONLINEARITIES

Nonlinearities can be classified as continuous and discontinuous. Becausediscontinuous nonlinearities cannot be locally approximated by linear functions, theyare also called "hard" nonlinearities. Hard nonlinearities are commonly found incontrol systems, both in small range operation and large range operation. Whether asystem in small range operation should be regarded as nonlinear or linear depends onthe magnitude of the hard nonlinearities and on the extent of their effects on thesystem performance.

Because of the common occurence of hard nonlinearities, let us briefly discussthe characteristics and effects of some important ones.

Saturation

When one increases the input to a physical device, the following phenomenon is oftenobserved: when the input is small, its increase leads to a corresponding (oftenproportional) increase of output; but when the input reaches a certain level, its further

=«*_

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (184)

170 Describing Function Analysis Chap. 5

increase does produces little or no increase of the output. The output simply staysaround its maximum value. The device is said to be in saturation when this happens.Simple examples are transistor amplifiers and magnetic amplifiers. A saturationnonlinearity is usually caused by limits on component size, properties of materials,and available power. A typical saturation nonlinearity is represented in Figure 5.8,where the thick line is the real nonlinearity and the thin line is an idealized saturationnonlinearity.

saturation- * 1

Linear

yw

/

1

/r_j

1 x

1

1 saturation1

Figure 5.8 : A saturation nonlinearity

Most actuators display saturation characteristics. For example, the outputtorque of a two-phase servo motor cannot increase infinitely and tends to saturate, dueto the properties of the magnetic material. Similarly, valve-controlled hydraulic servomotors are saturated by the maximum flow rate.

Saturation can have complicated effects on control system performance.Roughly speaking, the occurence of saturation amounts to reducing the gain of thedevice (e.g., the amplifier) as the input signals are increased. As a result, if a system isunstable in its linear range, its divergent behavior may be suppressed into a self-sustained oscillation, due to the inhibition created by the saturating component on thesystem signals. On the other hand, in a linearly stable system, saturation tends to slowdown the response of the system, because it reduces the effective gain.

On-off nonlinearity

An extreme case of saturation is the on-off or relay nonlinearity. It occurs when thelinearity range is shrunken to zero and the slope in the linearity range becomesvertical. Important examples of on-off nonlinearities include output torques of gas jetsfor spacecraft control (as in example 2.5) and, of course, electrical relays. On-offnonlinearities have effects similar to those of saturation nonlinearities. Furthermorethey can lead to "chattering" in physical systems due to their discontinuous nature.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (185)

Sect. 5.2

Dead-zone

Common Nonlinearities In Control Systems 171

In many physical devices, the output is zero until the magnitude of the input exceeds acertain value. Such an input-output relation is called a dead-zone. Consider forinstance a d.c. motor. In an idealistic model, we assume that any voltage applied tothe armature windings will cause the armature to rotate, with small voltage causingsmall motion. In reality, due to the static friction at the motor shaft, rotation willoccur only if the torque provided by the motor is sufficiently large. Similarly, whentransmitting motion by connected mechanical components, dead zones result frommanufacturing clearances. Similar dead-zone phenomena occur in valve-controlledpneumatic actuators and in hydraulic components.

-8

dead zoneFigure 5.9 : A dead-zone nonlinearity

Dead-zones can have a number of possible effects on control systems. Theirmost common effect is to decrease static output accuracy. They may also lead to limitcycles or system instability because of the lack of response in the dead zone. In somecases, however, they may actually stabilize a system or suppress self-oscillations. Forexample, if a dead-zone is incorporated into an ideal relay, it may lead to theavoidance of the oscillation at the contact point of the relay, thus eliminating sparksand reducing wear at the contact point. In chapter 8, we describe a dead-zonetechnique to improve the robustness of adaptive control systems with respect tomeasurement noise.

Backlash and hysteresis

Backlash often occurs in transmission systems. It is caused by the small gaps whichexist in transmission mechanisms. In gear trains, there always exist small gapsbetween a pair of mating gears, due to the unavoidable errors in manufacturing andassembly. Figure 5.10 illustrates a typical situation. As a result of the gaps, when thedriving gear rotates a smaller angle than the gap b, the driven gear does not move atall, which corresponds to the dead-zone (OA segment in Figure 5.10); after contacthas been established between the two gears, the driven gear follows the rotation of thedriving gear in a linear fashion (AB segment). When the driving gear rotates in thereverse direction by a distance of 2b, the driven gear again does not move,

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (186)

172 Describing Function Analysis Chap. 5

corresponding to the BC segment in Figure 5.10. After the contact between the twogears is re-established, the driven gear follows the rotation of the driving gear in thereverse direction (CD segment). Therefore, if the driving gear is in periodic motion,the driven gear will move in the fashion represented by the closed path EBCD. Notethat the height of B, C, D, E in this figure depends on the amplitude of the inputsinusoidal.

driven ;gear I

, drivinggear

outputangle

O A

b i n p u t

angle

Figure 5.10 : A backlash nonlinearity

A critical feature of backlash is its multi-valued nature. Corresponding to eachinput, two output values are possible. Which one of the two occur depends on thehistory of the input. We remark that a similar multi-valued nonlinearity is hysteresis,which is frequently observed in relay components.

Multi-valued nonlinearities like backlash and hysteresis usually lead to energystorage in the system. Energy storage is a frequent cause of instability and self-sustained oscillation.

5.3 Describing Functions of Common Nonlinearities

In this section, we shall compute the describing functions for a few commonnonlinearities. This will not only allow us to familiarize ourselves with the frequencydomain properties of these common nonlinearities, but also will provide furtherexamples of how to derive describing functions for nonlinear elements.

SATURATION

The input-output relationship for a saturation nonlinearity is plotted in Figure 5.11,with a and k denoting the range and slope of the linearity. Since this nonlinearity issingle-valued, we expect the describing function to be a real function of the input

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (187)

Sect. 5.3 Describing Functions of Common Nonlinearities 173

saturation

sinusoidalinput

w(t)unsaturatedoutput

saturatedoutput

CO/

Figure 5.11 : Saturation nonlinearity and the corresponding input-output relationship

amplitude.

Consider the input x(t) = Asin(otf)- If A < a, then the input remains in the linearrange, and therefore, the output is w(t) = kAsin(a)f). Hence, the describing function issimply a constant k.

Now consider the case A > a. The input and the output functions are plotted inFigure 5.11. The output is seen to be symmetric over the four quarters of a period. Inthe first quarter, it can be expressed as

w(t) =kA sin(cor)

ka <OH<7l/2

where y = sin ' (a/A). The odd nature of w(t) implies that a\=Q and the symmetry

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (188)

174 Describing Function Analysis

over the four quarters of a period implies that

4 rKl2

b 1 = - I vv(f) sin(G)0 d(& t)

4 r Y

Chap. 5

4 rY ~ A ft/2

= - kAsinz{(Ot)d((at)+-\ kasm(<£>t) d(($t)

2kA

(5.10)

Therefore, the describing function is

N(A)b\ 2k f . _i a a \, a2 ,

= = [sin » _ + _ 1 - — ]A A-

(5.11)

The normalized describing function (N(A)/k) is plotted in Figure 5.12 as afunction of A/a . One can observe three features for this describing function:

1. N(A) = kif the input amplitude is in the linearity range

2. N(A) decreases as the input amplitude increases

3. there is no phase shift

The first feature is obvious, because for small signals the saturation is not displayed.The second is intuitively reasonable, since saturation amounts to reduce the ratio ofthe output to input. The third is also understandable because saturation does not causethe delay of the response to input.

As a special case, one can obtain the describing function for the relay-type (on-off) nonlinearity shown in Figure 5.13. This case corresponds to shrinking the linearityrange in the saturation function to zero, i.e., a —> 0, k —> °°, but ka = M. Though b\can be obtained from (5.10) by taking the limit, it is more easily obtained directly as

4 i.n/2 4b, = _ I Msin(cof) d((Ot) = - M

n Jo ft

Therefore, the describing function of the relay nonlinearity is

N{A)=™L (5.12)

The normalized describing function (N/M) is plotted in Figure 5.13 as a function of

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (189)

Sect. 5.3 Describing Functions of Common Nonlinearities 175

N(A) 1 k

1.0 --

0.8 --

0.6 - •

0 . 4 - •

0.2-•

s— linearity range

t

! \i\! V! ^ - —

-4 1 1 ^

0 1. 5. 10. AlaFigure 5.12 : Describing function of thesaturation nonlinearity

input amplitude. Although the describing function again has no phase shift, the flatsegment seen in Figure 5.12 is missing in this plot, due to the completely nonlinearnature of the relay. The asymptic properties of the describing function curve in Figure5.13 are particularly interesting. When the input is infinitely small, the describingfunction is infinitely large. When the input is infinitely large, the describing functionis infinitely small. One can gain an intuitive understanding of these properties byconsidering the ratio of the output to input for the on-off nonlinearity.

\off

M

w

-M

on

X

N(A)1M

1.2-1.0-0.8-0.6-0.4 -0.2-

1 to infinity" \

\ \^ \- V

to zero ^

1 *-0 1. 5. 10. A

Figure 5.13 : Relay nonlinearity and its describing function

DEAD-ZONE

Consider the dead-zone characteristics shown in Figure 5.9, with the dead-zone widthbeing 25 and its slope k. The response corresponding to a sinusoidal inputx(t) = Asin(cor) into a dead-zone of width 28 and slope k, with A >5, is plotted inFigure 5.14. Since the characteristics is an odd function, al = 0. The response is alsoseen to be symmetric over the four quarters of a period. In one quarter of a period,i.e., when 0 < (at < n/2, one has

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (190)

176 Describing Function Analysis Chap. 5 t

Figure 5.14 : Input and output functions for a dead-zone nonlinearity

0 O<co/<y£(Asin(a>O-S) y<a>t< n/2

where y = sin ' (8/A). The coefficient bj can be computed as follows

4 ,rc/2 4 .71/2b\ = - w(t) sin(au) d{u>t) = - i(Asin(co/) - 8) sin(cor) d((Ot)

KJ0 KJy

(it 2

(5.13) |

This leads to

A A

This describing function /V(A) is a /-ea/ function and, therefore, there is no phase shift

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (191)

Sect. 5.3 Describing Functions of Common Nonlineariti.es 177

(reflecting the absence of time-delay). The normalized describing function is plotted inFigure 5.15. It is seen that N(A)/k is zero when A/8 < 1, and increases up to 1 withA/8. This increase indicates that the effect of the dead-zone gradually diminishes asthe amplitude of the input signal is increased, consistently with intuition.

Figure 5.15 : Describing function of thedead-zone nonlinearity

BACKLASH

The evaluation of the describing functions for backlash nonlinearity is more tedious.Figure 5.16 shows a backlash nonlinearity, with slope k and width 2b. If the inputamplitude is smaller than b, there is no output. In the following, let us consider theinput being x(t) = A sin(tt)f), A > b . The output w(t) of the nonlinearity is as shown inthe figure. In one cycle, the function w(t) can be represented as

w(t) = {A -b)k

w(t) = {A sin(co t) + b)k

w(t) = ~(A -b)k

w(t) = (/4sin(cor) -b)k

where y = s in ' 1 (1 - 2b/A).

n-

3TC

T

cor < 7t-y

<a>t<2n

~~2~

<oit<2n-y

<5n2

Unlike the previous nonlinearities, the function w{t) here is neither odd nor even.Therefore, «j and b\ are both nonzero. Using (5.7b) and (5.7c), we find through sometedious integrations that

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (192)

178 Describing Function Analysis Chap. 5

( 0 /

Figure 5.16 : Input and output functions for a backlash nonlinearity

4kb ,b , ,

1 jr. 2 A A i A

Therefore, the describing function of the backlash is given by

(5.14a)

(5.14b)

The amplitude of the describing function for backlash is plotted in Figure 5.17.

We note a few interesting points :

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (193)

Sect. 5.4 Describing Function A nalysis of Nonlinear Systems 179

0 0.2 0.4 0.6 0.8 1.0 JLA

Figure 5.17 : Amplitude of describingfunction for backlash

l.\N(A)\ = 0 ifA=b.

2. \N{A)\ increases, when b/A decreases.

3. |W(/4)|-> 1 as b/A-^ 0.

The phase angle of the describing function is plotted in Figure 5.18. Note that a phaselag (up to 90°) is introduced, unlike the previous nonlinearities. This phase lag is thereflection of the time delay of the backlash, which is due to the gap b. Of course, alarger b leads to a larger phase lag, which may create stability problems in feedbackcontrol systems.

IK

o

-20

-40

-60

-90 0 0.2 0.4 0.6 0.8 1.0 b_A

Figure 5.18 : Phase angle of describingfunction for backlash (degree)

5.4 Describing Function Analysis of Nonlinear Systems

For a nonlinear system containing a nonlinear element, we now know how to obtain adecribing function for the nonlinear element. The next step is to formalize theprocedure in subsection 5.1.1 for the prediction of limit cycles, based on the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (194)

180 Describing Function Analysis Chap. 5

describing function representation of the nonlinearity. The basic approach to achievethis is to apply an extended version of the famous Nyquist criterion in linear control tothe equivalent system. Let us begin with a short review of the Nyquist criterion andits extension.

5.4.1 The Nyquist Criterion and Its Extension

Consider the linear system of Figure 5.19. The characteristic equation of this system is

8(p) = 1 + G(p) H(p) = 0

Note that 8(p), often called the loop transfer function, is a rational function of p, withits zeros being the poles of the closed-loop system, and its poles being the poles of theopen-loop transfer function G(p) H(p). Let us rewrite the characteristic equation as

G(p) H(p) = -l

G(p)

H(p)Figure 5.19 : Closed-loop linear system

Based on this equation, the famous Nyquist criterion can be derived straightforwardlyfrom the Cauchy theorem in complex analysis. The criterion can be summarized(assuming that G(p) H(p) has no poles or zeros on the yco axis) in the followingprocedure (Figure 5.20):

1. draw, in the p plane, a so-called Nyquist path enclosing the right-halfplane

2. map this path into another complex plane through G(p)H{p)

3. determine /V, the number of clockwise encirclements of the plot ofG{p)H{p) around the point (- 1,0)

4. compute Z, the number of zeros of the loop transfer function 5(p) in theright-half p plane, by

Z = N + P , where P is the number of unstable poles of 5(p)

Then the value of Z is the number of unstable poles of the closed-loop system.

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (195)

Sect. 5.4 Describing Function Analysis of Nonlinear Systems 181

p plane

+ 00

G(p)H(p)

Nyquist path

-coco-> +co

Figure 5.20 : The Nyquist criterion

A simple formal extension of the Nyquist criterion can be made to the casewhen a constant gain K (possibly a complex number) is included in the forward pathin Figure 5.21. This modification will be useful in interpreting the stability analysis oflimit cycles using the describing function method. The loop transfer function becomes

8(p)=l+KG(p)H(p)

with the corresponding characteristic equation

G(p)H(p) = -

The same arguments as used in the derivation of Nyquist criterion suggest the sameprocedure for determining unstable closed-loop poles, with the minor difference thatnow Z represents the number of clockwise encirclements of the G(p) H(p) plot aroundthe point - l/K. Figure 5.21 shows the corresponding extended Nyquist plot.

Im

G(p) H(p)0

— • K

H(p)

G(p) —

Figure 5.21 : Extension of the Nyquist criterion

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (196)

182 Describing Function Analysis Chap. 5

5.4.2 Existence of Limit Cycles

Let us now assume that there exists a self-sustained oscillation of amplitude A andfrequency co in the system of Figure 5.22. Then the variables in the loop must satisfythe following relations:

x = -y

w = N(A,(o)x

(5.15)

(5.16)

Therefore, we have y = G(j(£>)N(A,(i))(-y). Because y * 0, this implies

GO'co) N(A,(0) + 1 = 0

which can be written as

G 0 " « » = - 1N(A,(o)

Therefore, the amplitude A and frequency CO of the limit cycles in the system mustsatisfy (5.16). If the above equation has no solutions, then the nonlinear system has nolimit cycles.

Expression (5.16) represents two nonlinear equations (the real part andimaginary part each give one equation) in the two variables A and CO. There areusually a finite number of solutions. It is generally very difficult to solve theseequations by analytical methods, particularly for high-order systems, and therefore, agraphical approach is usually taken. The idea is to plot both sides of (5.16) in thecomplex plane and find the intersection points of the two curves.

DescribingFunction Linear Element

r(t) = 0 x(t)

r *N(A, CO)

w(t)WCO)

y(t)

Figure 5.22 : A nonlinear system

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (197)

Sect. 5.4 Describing Function Analysis of Nonlinear Systems 183

FREQUENCY-INDEPENDENT DESCRIBING FUNCTION

First, we consider the simpler case when the describing function N being a function ofthe gain A only, i.e., N(A, co) = N(A). This includes all single-valued nonlinearitiesand some important double-valued nonlinearities such as backlash. The equalitybecomes

1N(A)

(5.17)

We can plot both the frequency response function G(j(o) (varying co) and the negativeinverse describing function (- l/N(A)) (varying A) in the complex plane, as in Figure5.23. If the two curves intersect, then there exist limit cycles, and the values of A andco corresponding to the intersection point are the solutions of Equation (5.17). If thecurves intersect n times, then the system has n possible limit cycles. Which one isactually reached depends on the initial conditions. In Figure 5.23, the two curvesintersect at one point K. This indicates that there is one limit cycle in the system. Theamplitude of the limit cycle is A^, the value of A corresponding to the point K on the- l/N(A) curve. The frequency of the limit cycle is co , the value of co correspondingto the point K on the G(jco) curve.

Figure 5.23 : Detection of limit cycles

Note that for single-valued nonlinearities, N is real and therefore the plot of- l/N always lies on the real axis. It is also useful to point out that, as we shalldiscuss later, the above procedure only gives a prediction of the existence of limitcycles. The validity and accuracy of this prediction should be confirmed by computersimulations.

We already saw in section 5.1.1 an example of the prediction of limit cycles, forthe Van der Pol equation.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (198)

184 Describing Function Analysis Chap. 5

FREQUENCY-DEPENDENT DESCRIBING FUNCTION

For the general case, where the describing function depends on both input amplitudeand frequency (N = N(A, co)), the method can be applied, but with more complexity.Now the right-hand side of (5.15), - l/N(A, co) , corresponds to a family of curves onthe complex plane with A as the running parameter and co fixed for each curve, asshown in Figure 5.24. There are generally an infinite number of intersection pointsbetween the G(y'co) curve and the - 1/N(A, co) curves. Only the intersection points withmatched co indicate limit cycles.

Im

Figure 5.24 : Limit cycle detection forfrequency-dependent describing functions

To avoid the complexity of matching frequencies at intersection points, it maybe advantageous to consider the graphical solution of (5.16) directly, based on theplots of G(ja>)N(A, co). With A fixed and co varying from 0 to °°, we obtain a curverepresenting G(J(o)N{A,(o). Different values of A correspond to a family of curves, asshown in Figure 5.25. A curve passing through the point (- 1,0) in the complex planeindicates the existence of a limit cycle, with the value of A for the curve being theamplitude of the limit cycle, and the value of co at the point (- 1,0) being thefrequency of the limit cycle. While this technique is much more straightforward thanthe previous one, it requires repetitive computation of the G(/co) in generating thefamily of curves, which may be handled easily by computer.

5.4.3 Stability of Limit Cycles

As pointed out in chapter 2, limit cycles can be stable or unstable. In the above, wehave discussed how to detect the existence of limit cycles. Let us now discuss how todetermine the stability of a limit cycle, based on the extended Nyquist criterion insection 5.4.1.

Consider the plots of frequency response and inverse describing function inFigure 5.26. There are two intersection points in the figure, predicting that the system

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (199)

Sect. 5.4 Describing Function Analysis of Nonlinear Systems 185

G(j(o)N(A,a>)Figure 5.25graphically

Solving equation (5.15)

has two limit cycles. Note that the value of A corresponding to point L[ is smallerthan the value of A corresponding to L2. For simplicity of discussion, we assume thatthe linear transfer function G(p) has no unstable poles.

Im

Figure 5.26 : Limit Cycle Stability

Let us first discuss the stability of the limit cycle at point Lj. Assume that thesystem initially operates at point Lj, with the limit cycle amplitude being A[, and itsfrequency being 00|. Due to a slight disturbance, the amplitude of the input to thenonlinear element is slightly increased, and the system operating point is moved fromLj to L | . Since the new point L] is encircled by the curve of G(ju>), according to theextended Nyquist criterion mentioned in section 5.4.1, the system at this operatingpoint is unstable, and the amplitudes of the system signals will increase. Therefore, theoperating point will continue to move along the curve - ]/N(A ) toward the other limitcycle point L2. On the other hand, if the system is disturbed so that the amplitude A asdecreased, with the operating point moved to the point Lj", then A will continue todecrease and the operating point moving away from Ll in the other direction. This is

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (200)

186 Describing Function Analysis Chap. 5

because L{ is not encircled by the curve G(yco) and thus the extended Nyquist plotasserts the stability of the system. The above discussion indicates that a slightdisturbance can destroy the oscillation at point Lj and, therefore, that this limit cycle isunstable. A similar analysis for the limit cycle at point L2 indicates that that limitcycle is stable.

Summarizing the above discussion and the result in the previous subsection, weobtain a criterion for existence and stability of limit cycles:

Limit Cycle Criterion: Each intersection point of the curve G(j(o) and the curve- l/N(A) corresponds to a limit cycle. If points near the intersection and along theincreasing-A side of the curve - 1 /N(A) are not encircled by the curve G(jai) , thenthe corresponding limit cycle is stable. Otherwise, the limit cycle is unstable.

5.4.4 Reliability of Describing Function Analysis

Empirical evidence over the last three decades, and later theoretical justification,indicate that the describing function method can effectively solve a large number ofpractical control problems involving limit cycles. However, due to the approximatenature of the technique, it is not surprising that the analysis results are sometimes notvery accurate. Three kinds of inaccuracies are possible:

1. The amplitude and frequency of the predicted limit cycle are notaccurate

2. A predicted limit cycle does not actually exist

3. An existing limit cycle is not predicted

The first kind of inaccuracy is quite common. Generally, the predictedamplitude and frequency of a limit cycle always deviate somewhat from the truevalues. How much the predicted values differ from the true values depends on howwell the nonlinear system satisfies the assumptions of the describing function method.In order to obtain accurate values of the predicted limit cycles, simulation of thenonlinear system is necessary.

The occurrence of the other two kinds of inaccuracy is less frequent but hasmore serious consequences. Usually, their occurrence can be detected by examiningthe linear element frequency response and the relative positions of the G plot and- l/.V plot.

\ 'iolation of filtering hypothesis: The validity of the describing function method relieson the filtering hypothesis defined by (5.5). For some linear elements, this hypothesis

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (201)

Sect. 5.5 Summary 187

is not satisfied and errors may result in the describing function analysis. Indeed, anumber of failed cases of describing function analysis occur in systems whose linearelement has resonant peaks in its frequency response G(/a>).

Graphical Conditions: If the G(yco) locus is tangent or almost tangent to the - l/Nlocus, then the conclusions from a describing function analysis might be erroneous.Such an example is shown in Figure 5.27(a). This is because the effects of neglectedhigher harmonics or system model uncertainty may cause the change of theintersection situations, particularly when filtering in the linear element is weak. As aresult, the second and third types of errors listed above may occur. A classic case ofthis problem involves a second-order servo with backlash studied by Nychols. Whiledescribing function analysis predicts two limit cycles (a stable one at high frequencyand an unstable one at low frequency), it can be shown that the low-frequencyunstable limit cycle does not exist.

Im Im

- 1 IN(A)

( a ) ( b )

Figure 5.27 : Reliability of limit cycle prediction

Conversely, if the - l/N locus intersects the G locus almost perpendicularly,then the results of the describing function are usually good. An example of thissituation is shown in Figure 5.27(b).

5.5 Summary

The describing function method is an extension of the frequency response method oflinear control. It can be used to approximately analyze and predict the behavior ofimportant classes of nonlinear systems, including systems with hard nonlinearities.The desirable properties it inherits from the frequency response method, such as its

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (202)

188 Describing Function Analysis Chap. 5

graphical nature and the physically intuitive insights it can provide, make it animportant tool for practicing engineers. Applications of the describing functionmethod to the prediction of limit cycles were detailed. Other applications, such aspredicting subharmonics, jump phenomena, and responses to external sinusoidalinputs, can be found in the literature.

5.6 Notes and References

An extensive and clear presentation of the describing function method can be found in [Gelb and

VanderVelde, 1968]. A more recent treatment is contained in [Hedrick, el al., 1982], which also

discusses specific applications to nonlinear physical systems. The describing function method was

developed and successfully used well before its mathematical justification was completely

formalized [Bergen and Franks, 1971]. Figures 5.14 and 5.16 are adapted from [Shinners, 1978]. The

Van der Pol oscillator example is adapted from [Hsu and Meyer, 1968].

5.7 Exercises

5.1 Determine whether the system in Figure 5.28 exhibits a self-sustained oscillation

cycle). If so, determine the stability, frequency, and amplitude of the oscillation.

limit

H)+1'

-1p(p +

K

>)(p + 2)

Figure 5.28 : A nonlinear system containing a relay

5.2 Determine whether the system in Figure 5.29 exhibits a self-sustained oscillation. If so,

determine the stability, frequency, and amplitude of the oscillation.

5.3 Consider the nonlinear system of Figure 5.30. Determine the largest K which preserves the

stability of the system. If K = 2Kmax, find the amplitude and frequency of the self-sustained

oscillation.

5.4 Consider the system of Figure 5.31, which is composed of a high-pass filter, a saturation

function, and the inverse low-pass filter. Show that the system can be viewed as a nonlinear low-

A

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (203)

Sect. 5.7 Exercises 189

Y = 0

p(p

Figure 5.29 : A nonlinear system

containing a dead-zone

pass filter, which attenuates high-frequency inputs without introducing a phase lag.

5.5 This exercise is based on a result of [Tsypkin, 1956].

Consider a nonlinear system whose output w(t) is related to the input u(t) by an odd function,

of the form

w(t) = F(u(t)) = - F( - u(t))

Derive the following very simple approximate formula for the describing function N(A)

N{A) = JL [ F(A) + F(AI2) ]

To this effect, you may want to use the fact that

(5.18)

fix) -dx = -6

where the remainder R verifies R = / 6 © / ( 2 ^ 6 ! ) for some £ € ( - 1 , 1 ) . Show that

approximation (5.18) is quite precise (how precise?).

o20

20

K

p(l+0.1p)(l+0.02p)

Figure 5.30 : A nonlinear system containing a saturation

* ^

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (204)

190 Describing Function Analysis Chap. S

v+1

v+1k-

_ / ^ i »a

x2p+l

V+i

Figure 5.31 : A nonlinear low-pass filter

Invert (5.18) so as to obtain for the input-output relation a solution of the form

5.6 In this exercise, adapted form [Phillips and Harbor, 1988], let us consider the system of Figure 31

5.32, which is typical of the dynamics of electronic oscillators used in laboratories, with

G(p) = -5p

+25

Use describing function analysis to predict whether the system exhibits a limit cycle, depending on jj

the value of the saturation level k. In such cases, determine the limit cycle's frequency and *

amplitude. %

Saturation Linear Element

w(t)G(p)

y(')

Figure 5.32 : Dynamics of an electronic oscillator

Interpret intuitively, by assuming that the system is started at some small initial state, and

noticing that y{t) can stay neither at small values (because of instability) nor at saturation values (by

applying the final value theorem of linear control).

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (205)

Part II

Nonlinear Control Systems Design

In Part I, we studied how to analyze the behavior of a nonlinear control system,assuming that the control system had been designed. Part II is devoted to the problemof designing nonlinear control systems. In this introduction, we discuss some generalissues involved in nonlinear control system design, particularly emphasizing thedifferences of nonlinear control design problems from linear ones. In the followingchapters, we will detail the specific control methods available to the designer.

As pointed out in chapter 1, the objective of control design can be stated asfollows: given a physical system to be controlled and the specifications of its desiredbehavior, construct a feedback control law to make the closed-loop system display thedesired behavior. In accordance with this design objective, we consider a number ofkey issues. First, two basic types of nonlinear control problems, nonlinear regulationand nonlinear tracking, are defined. Next, the specifications of the desired behavior ofnonlinear control systems are discussed. Basic issues in constructing nonlinearcontrollers are then outlined. Finally, the major methods available for designingnonlinear controllers are briefly surveyed.

191

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (206)

192 Nonlinear Control Systems Design Part II

II. 1 Nonlinear Control Problems

If the tasks of a control system involve large range and/or high speed motions,nonlinear effects will be significant in the dynamics and nonlinear control may benecessary to achieve the desired performance. Generally, the tasks of control systemscan be divided into two categories: stabilization (or regulation) and tracking (orservo). In stabilization problems, a control system, called a stabilizer (or a regulator),is to be designed so that the state of the closed-loop system will be stabilized aroundan equilibrium point. Examples of stabilization tasks are temperature control ofrefrigerators, altitude control of aircraft and position control of robot arms. In trackingcontrol problems, the design objective is to construct a controller, called a tracker, sothat the system output tracks a given time-varying trajectory. Problems such asmaking an aircraft fly along a specified path or making a robot hand draw straightlines or circles are typical tracking control tasks.

STABILIZATION PROBLEMS

In order to facilitate the analytic study of stabilization and tracking design in the laterchapters, let us provide some formal definitions of stabilization and tracking problems.

Asymptotic Stabilization Problem: Given a nonlinear dynamic system described by

x = f(x, u, t)

find a control law u such that, starting from anywhere in a region in Q., the state xtends toQ as t —> °o .

If the control law depends on the measurement signals directly, it is said to be astatic control law. If it depends on the measurement through a differential equation,the control law is said to be a dynamic control law, i.e., there is dynamics in thecontrol law. For example, in linear control, a proportional controller is a staticcontroller, while a lead-lag controller is a dynamic controller.

Note that, in the above definition, we allow the size of the region Q. to be large;otherwise, the stabilization problem may be adequately solved using linear control.Note also that if the objective of the control task is to drive the state to some non-zeroset-point xd, we can simply transform the problem into a zero-point regulationproblem by taking x - \d as the state.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (207)

Nonlinear Control Problems 193

(a) (b)

Figure II.l : (a) a pendulum; (b) an inverted pendulum, with cart

Example II.l: Stabilization of a pendulum

Consider the pendulum in Figure II. 1 (a). Its dynamics is

T (II.l)

Assume that our task is to bring the pendulum from a large initial angle, say 6(0) = 60° , to the

vertical-up position. One choice of the stabilizer is

) — k 6 — mglsinQ (H.2)

with kd and k denoting positive constants, This leads to the following globally stable closed-loop

dynamics

i.e., the controlled pendulum behaves as a stable mass-spring-damper system. Note that the

controller (II.2) is composed of a P.D. (proportional pius derivative) feedback part for stability

and a feedforward part for gravity compensation. Another interesting controller is

which leads to the stable closed-loop dynamics

This amounts to artificially reverting the gravity field and adding viscous damping.

This example illustrates the point that feedback and feedforward control actions amount to

modifying the dynamics of the plant into a desirable form. Lj

However, many nonlinear stabilization problems are not so easy to solve. Onesuch example is the inverted pendulum shown in Figure H.l(b) which can be easilyshown to have the following dynamics

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (208)

194 Nonlinear Control Systems Design Part II

(M + m)x+mlcosQQ-mlsmQQ2 = u (II.4a)

m'x cosG + ml 8 - mgsinQ = 0 (II.4b)

(where the mass of the cart is not assumed to be negligible). A particularly interestingtask is to design a controller to bring the inverted pendulum from a vertical-downposition at the middle of the lateral track to a vertical-up position at the same lateralpoint. This seeming simple nonlinear control problem is surprisingly difficult to solvein a systematic fashion (see Exercise II.5). This problem arises because there are twodegrees of freedom and only one input.

TRACKING PROBLEMS

The task of asymptotic tracking can be defined similarly.

Asymptotic Tracking Problem: Given a nonlinear dynamics system described by

x = f(x, u, t)

y = h(x)

and a desired output trajectory y^, find a control law for the input u such that,starting from any initial state in a region O, the tracking errors y(t) - y^f) go to zero,while the whole state x remains bounded.

Note that, from a practical point of view, one may require that x actually remain"reasonably" bounded, and, in particular, within the range of validity of the systemmodel. This may be verified either analytically, or in simulations.

When the closed-loop system is such that proper initial states imply zerotracking error for all the time,

y(0 = y<fit) Vr>o

the control system is said to be capable of perfect tracking. Asymptotic trackingimplies that perfect tracking is asymptotically achieved. Exponential trackingconvergence can be defined similarly.

Throughout the rest of the book, unless otherwise specified, we shall make themild assumption that the desired trajectory y^ and its derivatives up to a sufficientlyhigh order (generally equal to the system's order) are continuous and bounded. Wealso assume that yj(t) and its derivatives are available for on-line control computation.This latter assumption is satisfied by control tasks where the desired output yd{i) isplanned ahead of time. For example, in robot tracking tasks, the desired positionhistory is generally planned ahead of time and its derivatives can be easily obtained.

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (209)

Nonlinear Control Problems 195

Actually, smooth time-histories are often generated themselves through a filteringprocess, thereby automatically providing higher derivatives of the desired output. Insome tracking tasks, however, the assumption is not satisfied, and a so-calledreference model may be used to provide the required derivative signals. For example,in designing a tracking control system for the antenna of a radar so that it will closelypoint toward an aircraft at all times, we only have the position of the aircraft ya{f)available at a given time instant (assuming that it is too noisy to be numericallydifferentiated). However, generally the tracking control law will also use thederivatives of the signals to be tracked. To solve this problem, we can generate thedesired position, velocity and acceleration to be tracked by the antenna using thefollowing second-order dynamics

y'd+k\y<i+hyd = hyjfi (IL5>

where kl and £2 a r e chosen positive constants. Thus the problem of following the

aircraft is translated into the problem of tracking the output yjjt) of the referencemodel. Note that the reference model serves the dual purpose of providing the desiredoutput of the tracking system in response to the aircraft position measurements, andgenerating the derivatives of the desired output for tracker design. Of course, for theapproach to be effective, the filtering process described by (II.5) should be fast enoughfor y^t) to closely approximate ya{t).

For non-minimum phase systems (precise definitions of nonlinear non-minimum phase systems will be provided in chapter 6), perfect tracking andasymptotic tracking cannot be achieved, as seen in the following example.

Example II.2: Tracking control of a non-minimum phase linear system

Consider the linear system

y + 2y + 2y = — it + u

The system is non-minimum phase because it has a zero at p = 1. Assume that perfect tracking is

achieved, i.e., that y(l) = }></('). V ? > 0. Then, the input u satisfies

u-u = -(yd+2yd + 2yd)

Since this represents an unstable dynamics, u diverges exponentially. Note that the above

dynamics has a pole which exactly coincides with the unstable zero of the original system, i.e.,

perfect tracking for non-minimum phase systems can be achieved only by infinite control inputs.

By writing u as

p2 + 2n + 2

" = " P - \ yd

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (210)

196 Nonlinear Control Systems Design Part II

we see that the perfect-tracking controller is actually inverting the plant dynamics. C]

The inability of perfect tracking for a non-minimum phase linear system has itsroots in its inherent tendency of "undershooting" in its step response. Thus, the controldesign objective for non-minimum phase systems should not be perfect tracking orasymptotic tracking. Instead, we should be satisfied with, for example, bounded-errortracking, with small tracking error being achieved for desired trajectories of particularinterest.

RELATIONS BETWEEN STABILIZATION AND TRACKING PROBLEMS

Normally, tracking problems are more difficult to solve than stabilizationproblems, because in tracking problems the controller should not only keep the wholestate stabilized but also drive the system output toward the desired output. However,from a theoretical point of view, tracking design and stabilization design are oftenrelated. For instance, if we are to design a tracker for the plant

y+f(y,y,u) = o

so that e(t) = y(t) — y^t) goes to zero, the problem is equivalent to the asymptoticstabilization of the system

e+f(e,e,u,yd,yd,yd) = O (II.6)

whose state components are e and e. Clearly, the tracker design problem is solved ifwe know how to design a stabilizer for the non-autonomous dynamics (II.6).

On the other hand, stabilization problems can often be regarded as a specialcase of tracking problems, with the desired trajectory being a constant. In modelreference control, for instance, a set-point regulation problem is transformed into atracking problem by incorporating a reference model to filter the supplied set-pointvalue and generate a time-varying output as the ideal response for the tracking controlsystem.

II.2 Specifying the Desired Behavior

In linear control, the desired behavior of a control system can be systematicallyspecified, either in the time-domain (in terms of rise time, overshoot and settling timecorresponding to a step command) or in the frequency domain (in terms of regions inwhich the loop transfer function must lie at low frequencies and at high frequencies).In linear control design, one first lays down the quantitative specifications of theclosed-loop control system, and then synthesizes a controller which meets these

A

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (211)

Specifying the Desired Behavior 197

specifications. However, systematic specification for nonlinear systems (except thoseequivalent to linear systems) is much less obvious because the response of a nonlinearsystem to one command does not reflect its response to another command, andfurthermore a frequency-domain description is not possible.

As a result, for nonlinear systems, one often looks instead for some qualitativespecifications of the desired behavior in the operating region of interest. Computersimulation is an important complement to analytical tools in determining whether suchspecifications are met. Regarding the desired behavior of nonlinear control systems, adesigner can consider the following characteristics:

Stability must be guaranteed for the nominal model (the model used fordesign), either in a local sense or in a global sense. The region of stability andconvergence are also of interest.

Accuracy and speed of response may be considered for some "typical" motiontrajectories in the region of operation. For some classes of systems, appropriatecontroller design can actually guarantee consistent tracking accuracy independently ofthe desired trajectory, as discussed in chapter 7.

Robustness is the sensitivity to effects which are not considered in the design,such as disturbances, measurement noise, unmodeled dynamics, etc. The systemshould be able to withstand these neglected effects when performing the tasks ofinterest.

Cost of a control system is determined mainly by the number and type ofactuators, sensors, and computers necessary to implement it. The actuators, sensorsand the controller complexity (affecting computing requirement) should be chosenconsistently and suit the particular application.

A couple of remarks can be made at this point. First, stability does not imply theability to withstand persistent disturbances of even small magnitude (as discussed insection 4.9.2). The reason is that stability of a nonlinear system is defined with respectto initial conditions, and only temporary disturbances may be translated as initialconditions. For example, a stable control system may guarantee an aircraft's ability towithstand gusts, while being inept at handling windshears even of small magnitude.This situation is different from that of linear control, where stability always impliesability to withstand bounded disturbances (assuming that the system does stay in itslinear range). The effects of persistent disturbance on nonlinear system behavior areaddressed by the concept of robustness. Secondly, the above qualities conflict to someextent, and a good control system can be obtained only based on effective trade-offs interms of stability/robustness, stability/performance, cost/performance, and so on.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (212)

198 Nonlinear Control Systems Design Part II

II.3 Some Issues In Constructing Nonlinear Controllers

We now briefly describe a few aspects of controller design.

A Procedure for Control Design

Given a physical system to be controlled, one typically goes through the followingstandard procedure, possibly with a few iterations:

1. specify the desired behavior, and select actuators and sensors;

2. model the physical plant by a set of differential equations;

3. design a control law for the system;

4. analyze and simulate the resulting control system;

5. implement the control system in hardware.

Experience, creativity, and engineering judgment are all important in this process. Oneshould consider integrating control design with plant design if possible, similarly tothe many advances in aircraft design which have been achieved using linear controltechniques. Sometimes the addition or relocation of actuators and sensors may makean otherwise intractable nonlinear control problem easy.

Modeling Nonlinear Systems

Modeling is basically the process of constructing a mathematical description (usuallya set of differential equations) for the physical system to be controlled. Two points canbe made about modeling. First, one should use good understanding of the systemdynamics and the control tasks to obtain a tractable yet accurate model for controldesign. Note that more accurate models are not always better, because they mayrequire unnecessarily complex control design and analysis and more demandingcomputation. The key here is to keep "essential" effects and discard insignificanteffects in the system dynamics in the operating range of interest. Second, modeling ismore than obtaining a nominal model for the physical system: it should also providesome characterization of the model uncertainties, which may be used for robustdesign, adaptive design, or merely simulation.

Model uncertainties are the differences between the model and the real physicalsystem. Uncertainties in parameters are called parametric uncertainties while theothers are called non-parametric uncertainties. For example, for the model of acontrolled mass

m'x = u

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (213)

Some Issues In Constructing Nonlinear Controllers 199

the uncertainty in m is parametric uncertainty, while the neglected motor dynamics,measurement noise, sensor dynamics are non-parametric uncertainties. Parametricuncertainties are often easy to characterize. For example, m may be known to liesomewhere between 2 kg and 5 kg. The characterization of unmodeled dynamics fornonlinear systems is often more difficult, unlike the linear control case wherefrequency-domain characterizations can be systematically applied.

Feedback and Feedforward

In nonlinear control, the concept of feedback plays a fundamental role in controllerdesign, as it does in linear control. However, the importance of feedforward is muchmore conspicuous than in linear control. Feedforward is used to cancel the effects ofknown disturbances and provide anticipative actions in tracking tasks. Very often it isimpossible to control a nonlinear system stably without incorporating feedforwardaction in the control law. Note that a model of the plant is always required forfeedforward compensation (although the model need not be very accurate).

Asymptotic tracking control always requires feedforward actions to provide theforces necessary to make the required motion. It is interesting to note that manytracking controllers can be written in the form

u = feedforward + feedback

or in a similar form. The feedforward part intends to provide the necessary input forfollowing the specified motion trajectory and canceling the effects of the knowndisturbances. The feedback part then stabilizes the tracking error dynamics.

As an illustration of the use of feedforward, let us consider tracking controllerdesign in the familiar context of linear systems (as applicable to devices such as an x-yplotter, for instance). The discussion is interesting in its own right, since tracking oftime-varying trajectories is not commonly emphasized in linear control texts.

Example H.3: Tracking control of linear systems

Consider a linear (controllable and observable) minimum-phase system in the form

A{p)y = B(p)u (II.7)

where

ao + alp + .... + an_lp"-i + p"

The control objective is to make the output y(t) follow a time-varying desired trajectory yjit). We

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (214)

200 Nonlinear Control Systems Design Part II

assume that only the output y(t) is measured, and that y^, yd, ...,yj-r^ are known, with r being the

relative degree (the excess of poles over zeros) of the transfer function (thus, r=n-m).

The control design can be achieved in two steps. First, let us take the control law in the form

of

where v is a new input to be determined. Substitution of (II.8) into (II.7) leads to

A(p)e = B(p)v (II.9)

where e(t) = y(t) - yjf) is the tracking error. The feedforward signal (AIB) yd can be computed as

A - (r)

where the a,- ( i = 1,... , r) are constants obtained from dividing A by B, and w(t) is a filtered

version of yjj).

The second step is to construct input u so that the error dynamics is asymptotically stable.

Since e is known (by subtracting the known yd from the measured y), while its derivatives are not,

one can stabilize e by using standard linear techniques, e.g., pole-placement together with a

Luenberger observer. A simpler way of deriving the control law is to let

(11.10)D(p)

with C and D being polynomials of order (n-m). With this control law, the closed loop dynamics

is

If the coefficients of C and D are chosen properly, the poles of the closed-loop polynomial can be

placed anywhere in the complex plane (with the complex poles in conjugate pairs), as we shall

see in chapter 8. Therefore, the control law

" = 5 ^ + §< (IL11)

will guarantee that the tracking error e(t) remains at zero if initial conditions satisfy

yl'\0) = yJ-'XQ) (i = 1, ... , ; ) , and exponentially converges to zero if the initial conditions do not

satisfy these conditions.

The block diagram of the closed-loop system is depicted in Figure II.3. We can make the

following comments about the control system:

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (215)

Some Issues In Constructing Nonlinear Controllers 201

~ 1

A

B

CD

BA

y

Figure II.2 : Linear Tracking Control system

• The feedforward part of the control law, computed by inverting the plant model, is

responsible for reducing and eliminating the tracking errors, while the feedback part

results in stability of the whole system. If some derivatives of the desired trajectories

yj,t) are not available, one can simply omit them from the feedforward, which will only

cause bounded error in tracking. Note that one may easily adapt the above controller for

model reference control, with yd and its derivatives provided by the reference model.

• The control law (11.10) is equivalent to implementing a reduced-order Luenberger

observer. Higher order observers can also be used, possibly for the purpose of increasing

system robustness by exploiting the added flexibility.

• The above method cannot be directly used for tracking control of non-minimum phase

systems (with some of the roots of B(p) having positive real parts) since the inverse

model A/B is unstable. However, by feedforwarding low-frequency components of the

desired trajectories, good tracking in the low-frequency range (lower than the left-half

plane zeros of the plant) may still be achieved. For instance, by using (A/B{) yd as the

feedforward signal, the tracking error can be easily found to be

e(t) =C B

AC+BDLB

If B| is close to B at low frequencies, the control system can track a slowly varying yj{i)

well. One particular choice of 6 | is to eliminate the right half-plane zeros of B, which

will lead to good tracking for desired trajectories with frequencies lower than the right

half-plane zeroes. C]

Importance of Physical Properties

In linear control, it is common practice to generate a set of differential equations for aphysical system, and then forget where they came from. This presents no majorproblem there, at least in a theoretical sense, because linear control theory provides

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (216)

202 Nonlinear Control Systems Design Part II

powerful tools for analysis and design. However, such a procedure is typicallyundesirable for nonlinear systems, because the number of tools available for attackingnonlinear problems is comparatively limited. In nonlinear control design, exploitationof the physical properties can sometimes make control design for complex nonlinearplants a simple issue, or may easily solve an otherwise intractable design problem.This point is forcibly demonstrated in the solution of the adaptive robot controlproblem. Adaptive control of robot manipulators was long recognized to be far out ofthe reach of conventional adaptive control theory, because a robot's dynamics isstrongly nonlinear and has multiple inputs. However, the use of two physical facts,namely, the positive definiteness of the inertia matrix and the possibility of linearlyparametrizing robot dynamics, successfully led to an adaptive controller with thedesirable properties of global stability and tracking convergence, as shown inchapter 9.

Discrete Implementations

As discussed in chapter 1, nonlinear physical systems are continuous in nature and arehard to meaningfully discretize, while digital control systems may be treated ascontinuous-time systems in analysis and design if high sampling rates are used(specific quantifications are discussed, e.g., in section 7.3). Thus, we performnonlinear system analysis and controller design in continuous-time form. However, ofcourse, the control law is generally implemented digitally.

Numerical integration and differentiation are sometimes explicit parts of acontroller design. Numerical differentiation may avoid the complexity of constructingthe whole system state based on partial measurements (the nonlinear observerproblem), while numerical integration is a standard component of most adaptivecontroller designs, and can also be needed more generally in dynamic controllers.

Numerical differentiation may be performed in many ways, all aimed at gettinga reasonable estimate of the time-derivative, while at the same time avoiding thegeneration of large amounts of noise. One can use, for instance, a filtereddifferentiation of the form

p+a p+a

where p is the Laplace variable and a » 1 . The discrete implementation of theabove equation, assuming e.g., a zero-order hold, is simply

where the constants a j and a2 a r e defined as

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (217)

Available Methods of Nonlinear Control Design 203

ax = e~aT a2 = \ - ax

and T is the sampling period. Note that this approximate procedure can actually beinterpreted as building a reduced-order observer for the system. However, it does notuse an explicit model, so that the system can be nonlinear and its parameters unknown.An alternative choice of filter structure is studied in Exercise II.6.

Numerical integration actually consists in simulating in real-time some(generally nonlinear) dynamic subcomponent required by the controller. Manymethods can again be used. The simplest, which can work well at high sampling ratesor for low-dimensional systems, is the so-called Euler integration

xT

where T is the sampling period. A more sophisticated approach, which is veryeffective in most cases but is more computationally involved than mere Eulerintegration, is two-step Adams-Bashforth integration

x(t) - x(t-T) + (-x(t-T)--x(t-2T))T

More complex techniques may also be used, depending on the desired trade-offbetween accuracy and on-line computational efficiency.

II.4 Available Methods of Nonlinear Control Design

As in the analysis of nonlinear control systems, there is no general method fordesigning nonlinear controllers. What we have is a rich collection of alternative andcomplementary techniques, each best applicable to particular classes of nonlinearcontrol problems.

Trial-and-error

Based on the analysis methods provided in Part I, one can use trial-and-error tosynthesize controllers, similarly to, e.g., linear lead-lag controller design based onBode plots. The idea is to use the analysis tools to guide the search for a controllerwhich can then be justified by analysis and simulations. The phase plane method, thedescribing function method, and Lyapunov analysis can all be used for this purpose.Experience and intuition are critical in this process. However, for complex systemstrial-and-error often fails.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (218)

204 Nonlinear Control Systems Design Part II

Feedback linearization

As discussed earlier, the first step in designing a control system for a given physicalplant is to derive a meaningful model of the plant, i.e., a model that captures the keydynamics of the plant in the operational range of interest. Models of physical systemscome in various forms, depending on the modeling approach and assumptions. Someforms, however, lend themselves more easily to controller design. Feedbacklinearization deals with techniques for transforming original system models intoequivalent models of a simpler form.

Feedback linearization can be used as a nonlinear design methodology. Thebasic idea is to first transform a nonlinear system into a (fully or partially) linearsystem, and then use the well-known and powerful linear design techniques tocomplete the control design. The approach has been used to solve a number ofpractical nonlinear control problems. It applies to important classes of nonlinearsystems (so-called input-state linearizable or minimum-phase systems), and typicallyrequires full state measurement. However, it does not guarantee robustness in the faceof parameter uncertainty or disturbances.

Feedback linearization techniques can also be used as model-simplifyingdevices for robust or adaptive controllers, to be discussed next.

Robust control

In pure model-based nonlinear control (such as the basic feedback linearizationcontrol approach), the control law is designed based on a nominal model of thephysical system. How the control system will behave in the presence of modeluncertainties is not clear at the design stage. In robust nonlinear control (such as, e.g.,sliding control), on the other hand, the controller is designed based on theconsideration of both the the nominal model and some characterization of the modeluncertainties (such as the knowledge that the load to be picked up and carried by arobot is between 2 kg and 10 kg). Robust nonlinear control techniques have provenvery effective in a variety of practical control problems. They apply best to specificclasses of nonlinear systems, and generally require state measurements.

Adaptive control

Adaptive control is an approach to dealing with uncertain systems or time-varyingsystems. Although the term "adaptive" can have broad meanings, current adaptivecontrol designs apply mainly to systems with known dynamic structure, but unknownconstant or slowly-varying parameters. Adaptive controllers, whether developed forlinear systems or for nonlinear systems, are inherently nonlinear.

A

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (219)

Exercises 205

Systematic theories exist for the adaptive control of linear systems. Existingadaptive control techniques can also treat important classes of nonlinear systems, withmeasurable states and linearly parametrizable dynamics. For these nonlinear systems,adaptive control can be viewed as an alternative and complementary approach torobust nonlinear control techniques, with which it can be combined effectively.Although most adaptive control results are for single-input single-output systems,some important nonlinear physical systems with multiple-inputs have also beenstudied successfully.

Gain-scheduling

Gain scheduling (see [Rugh, 1991] for a recent discussion, and references therein) isan attempt to apply the well developed linear control methodology to the control ofnonlinear systems. It was originally developed for the trajectory control of aircraft.The idea of gain-scheduling is to select a number of operating points which cover therange of the system operation. Then, at each of these points, the designer makes alinear time-invariant approximation to the plant dynamics and designs a linearcontroller for each linearized plant. Between operating points, the parameters of thecompensators are then interpolated, or scheduled, thus resulting in a globalcompensator. Gain scheduling is conceptually simple, and, indeed, practicallysuccessful for a number of applications. The main problem with gain scheduling isthat has only limited theoretical guarantees of stability in nonlinear operation, but usessome loose practical guidelines such as "the scheduling variables should changeslowly" and "the scheduling variables should capture the plant's nonlinearities".Another problem is the computational burden involved in a gain-scheduling design,due to the necessity of computing many linear controllers.

II.5 Exercises

II. 1 Why do linear systems with a right half-plane zero exhibit the so-called "undershooting"

phenomenon (the step response initially goes downward)? Is the inability of perfect tracking for non-

minimum phase systems related to the undershooting phenomenon?

Consider for instance the system

Sketch its step response and compare it with that of the system

p1 + 2p+2

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (220)

206 Nonlinear Control Systems Design Part II

What are the differences in frequency responses?

What does the step response of a non-minimum phase linear system look like if it has two

right half-plane zeros? Interpret and comment.

11.2 Assume that you are given a pendulum and the task of designing a control system to track the

desired trajectory

6d(/) = /4sinco» 0<A<9Q" 0<cc<10tfz

What hardware components do you need to implement the control system? What requirements does

the task impose on the the specifications of the components? Provide a detailed outline of your

control system design.

11.3 List the model uncertainties associated with the pendulum model (II. 1). Discuss how to

characterize them.

11.4 Carry out the tracking design for the linear plants

Simulate their responses to the desired trajectories

with (O being 0.5, 1.5, and 4. rad/sec.

11.5 Figure out an energy-based strategy to bring the inverted pendulum in Figure Il.l.b from the

vertical-down position to the vertical-up position. (Hint: You may want first to express the system's

kinetic energy in a modified coordinate system chosen such that rotation and translation are

uncoupled.)

What does your controller guarantee along the x direction? Does it reduce to the usual linear

inverted pendulum controller when linearized?

11.6 An alternative to the filtered differentiation (11.12) consists in simply passing an approximate

derivative through a zero-order hold discrete filter, e.g.,

c\xold

where c, = e~aT . Discuss the relative merits of the two approaches.

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (221)

Chapter 6Feedback Linearization

Feedback linearization is an approach to nonlinear control design which has attracted agreat deal of research interest in recent years. The central idea of the approach is toalgebraically transform a nonlinear system dynamics into a (fully or partly) linear one,so that linear control techniques can be applied. This differs entirely fromconventional linearization (i.e., Jacobian linearization, as in section 3.3) in thatfeedback linearization is achieved by exact state transformations and feedback, ratherthan by linear approximations of the dynamics.

The idea of simplifying the form of a system's dynamics by choosing a differentstate representation is not entirely unfamiliar. In mechanics, for instance, it is wellknown that the form and complexity of a system model depend considerably on thechoice of reference frames or coordinate systems. Feedback linearization techniquescan be viewed as ways of transforming original system models into equivalent modelsof a simpler form. Thus, they can also be used in the development of robust oradaptive nonlinear controllers, as discussed in chapters 7 and 8.

Feedback linearization has been used successfully to address some practicalcontrol problems. These include the control of helicopters, high performance aircraft,industrial robots, and biomedical devices. More applications of the methodology arebeing developed in industry. However, there are also a number of importantshortcomings and limitations associated with the feedback linearization approach.Such problems are still very much topics of current research.

207

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (222)

208 Feedback Linearization Chap. 6

This chapter provides a description of feedback linearization, including what itis, how to use it for control design and what its limitations are. In section 6.1, thebasic concepts of feedback linearization are described intuitively and illustrated withsimple examples. Section 6.2 introduces mathematical tools from differentialgeometry which are useful to generalize these concepts to a broad class of nonlinearsystems. Sections 6.3 and 6.4 describe feedback linearization theory for SISO systems,and section 6.5 extends the methodology to MIMO systems.

6.1 Intuitive Concepts

This section describes the basic concepts of feedback linearization intuitively, usingsimple examples. The following sections will formalize these concepts for moregeneral nonlinear systems.

6.1.1 Feedback Linearization And The Canonical Form

In its simplest form, feedback linearization amounts to canceling the nonlinearities ina nonlinear system so that the closed-loop dynamics is in a linear form. This verysimple idea is demonstrated in the following example.

Example 6.1: Controlling the fluid level in a tank

Consider the control of the level h of fluid in a tank (Figure 6.1) to a specified level hd. The

control input is the flow u into the tank, and the initial level is ho.

outputflow

Figure 6.1 : Fluid level control in a tank

The dynamic model of the tank is

(6.1)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (223)

Sect. 6.1 Intuitive Concepts 209

where A(h) is the cross section of the tank and a is the cross section of the outlet pipe. If the

initial level ha is quite different from the desired level hd, the control of h involves a nonlinear

regulation problem.

The dynamics (6.1) can be rewritten as

A(h)'h = u-a-\lgh

If «(/) is chosen as

u(t) = a^2gh +A(h)v (6.2)

with v being an "equivalent input" to be specified, the resulting dynamics is linear

A = v

Choosing v as

v = -ah (6.3)

with h = h(t) - hd being the level error, and a being a strictly positive constant, the resulting

closed loop dynamics is

h + ah = 0 (6.4)

This implies that h(t) —> 0 as t —* °°. Based on (6.2) and (6.3), the actual input flow is

determined by the nonlinear control law

u(t) = a^2gh -A(h)ah (6.5)

Note that, in the control law (6.5), the first part on the right-hand side is used to provide the

output flow a~^2gh , while the second part is used to raise the fluid level according to the the

desired linear dynamics (6.4).

Similarly, if the desired level is a known time-varying function hd(t), the equivalent input v

can be chosen as

v = hd(t)-ah

so as to still yield H(t) -> 0 as / -» °° . •

The idea of feedback linearization, i.e., of canceling the nonlinearities andimposing a desired linear dynamics, can be simply applied to a class of nonlinearsystems described by the so-called companion form, or controllability canonical form.A system is said to be in companion form if its dynamics is represented by

xW =/(x) + b(x) u (6.6)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (224)

210 Feedback Linearization Chap. 6

where u is the scalar control input, x is the scalar output of interest,x = [x,x, ... ,x(n~l}]T is the state vector, and /(x) and b(x) are nonlinearfunctions of the states. This form is unique in the fact that, although derivatives of xappear in this equation, no derivative of the input u is present. Note that, in state-spacerepresentation, (6.6) can be written

X2

f(x) + b(x)u

For systems which can be expressed in the controllability canonical form, usingthe control input (assuming b to be non-zero)

ddf

x\

xn-\

u=l[v-f] (6.7)

we can cancel the nonlinearities and obtain the simple input-output relation (multiple-integrator form)

x(n) = v

Thus, the control law

with the ki chosen so that the polynomial pn + kn_\p"~l + .... + k0 has all its rootsstrictly in the left-half complex plane, leads to the exponentially stable dynamics

which implies that x(t) —> 0. For tasks involving the tracking of a desired output Xj(t),the control law

v = xdW - koe - k2e - .... - kn_{ e^-V (6.8)

(where e{t) = x(t) - xd{t) is the tracking error) leads to exponentially convergenttracking. Note that similar results would be obtained if the scalar x was replaced by avector and the scalar b by an invertible square matrix.

One interesting application of the above control design idea is in robotics. Thefollowing example studies control design for a two-link robot. Design for moregeneral robots is similar and will be discussed in chapter 9.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (225)

Sect. 6.1

Example 6.2: Feedback linearization of a two-link robot

Intuitive Concepts 211

Figure 6.2 provides the physical model of a two-link robot, with each joint equipped with a motor

for providing input torque, an encoder for measuring joint position, and a tachometer for

measuring joint velocity. The objective of the control design is to make the joint positions q s and

q2 follow desired position histories q^(t) and q^t) , which are specified by the motion planning

system of the robot. Such tracking control problems arise when a robot hand is required to move

along a specified path, e.g., to draw circles.

Figure 6.2 : A two-link robot

Using the well-known Lagrangian equations in classical dynamics, one can easily show that

the dynamic equations of the robot is

(6.9)1

"21 "22|52_J_

~hq2

hqx

-hqx-hq2

LS\

82

= [<7j q2]T being the two joint angles, T = [lj X 2 ] r being the joint inputs, and

Hu=mxl , 2 + / 2 + 2 / , / co S < ? 2 ]+ / 2

I = H2\ = m2l\ ' c ^ ^ + m2lc22 + !2

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (226)

212 Feedback Linearization Chap. 6

Equation (6.9) can be compactly expressed as

H(q)q + C(q, q)q + g(q) = x

with H, C and g defined obviously. Thus, by multiplying both sides by H ~ ' (the invertibility of

H is a physical property of the system, as discussed in Chapter 9), the above vector equation can

be put easily in the form of (6.6), with n = 2, although this dynamics now involves multiple inputs

and multiple outputs.

To achieve tracking control tasks, one can use the following control law

»2I»22»2

-hq2 -h(6.10)

where

with v = [ v | i>2 ] r being the equivalent input, q = q — q being the position tracking error and X

a positive number. The tracking error q then satisfies the equation

and therefore converges to zero exponentially. The control law (6.10) is commonly referred to as

"computed torque" control in robotics. It can be applied to robots with arbitrary numbers of joints,

as discussed in chapter 9. D

Note that in (6.6) we have assumed that the dynamics is linear in terms of thecontrol input u (although nonlinear in the states). However, the approach can be easilyextended to the case when u is replaced by an invertible function g(u). For example, insystems involving flow control by a valve, the dynamics may be dependent on M4

rather than directly on u, with u being the valve opening diameter. Then, by definingw = M4 , one can first design w similarly to the previous procedure and then computethe input « by u = w1^4 . This means that the nonlinearity is simply undone in thecontrol computation.

When the nonlinear dynamics is not in a controllability canonical form, onemay have to use algebraic transformations to first put the dynamics into thecontrollability form before using the above feedback linearization design, or to rely onpartial linearization of the original dynamics, instead of full linearization. These are

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (227)

Sect. 6.1 Intuitive Concepts 213

the topics of the next subsections. Conceptually, such transformations are not totallyunfamiliar: even in the case of linear systems, pole placement is often most easilyachieved by first putting the system in the controllability canonical form.

6.1.2 Input-State Linearization

Consider the problem of designing the control input u for a single-input nonlinearsystem of the form

x = f (x, u)

The technique of input-state linearization solves this problem in two steps. First, onefinds a state transformation z = z(x) and an input transformation u = u(x, v) so thatthe nonlinear system dynamics is transformed into an equivalent linear time-invariantdynamics, in the familiar form z = Az + bv. Second, one uses standard lineartechniques (such as pole placement) to design v.

Let us illustrate the approach on a simple second-order example. Consider thesystem

i | = - 2x\ + <xx2 + sin xx (6.11a)

x2 = - x 2 cosx[ + M cos(2x() (6.11b)

Even though linear control design can stabilize the system in a small region around theequilibrium point (0, 0), it is not obvious at all what controller can stabilize it in alarger region. A specific difficulty is the nonlinearity in the first equation, whichcannot be directly canceled by the control input u.

However, if we consider the new set of state variables

zx=xx (6.12a)

Z2 = ax2+ sinxj (6.12b)

then, the new state equations are

z 1 = - 2 z 1 + z 2 (6.13a)

z2 = - 2zj cos Zj + coszj sinzj + au cos(2z[) (6.13b)

Note that the new state equations also have an equilibrium point at (0, 0). Now we seethat the nonlinearities can be canceled by the control law of the form

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (228)

214 Feedback Linearization Chap. 6

u = (v-cosz j sin Zj + 2zj cos Zj) (6.14)

where v is an equivalent input to be designed (equivalent in the sense that determiningv amounts to determining «, and vice versa), leading to a linear input-state relation

z l = - 2 z 1 + z 2 (6.15a)

z2 = v (6.15b)

Thus, through the state transformation (6.12) and input transformation (6.14), theproblem of stabilizing the original nonlinear dynamics (6.11) using the original controlinput u has been transformed into the problem of stabilizing the new dynamics (6.15)using the new input v.

Since the new dynamics is linear and controllable, it is well known that thelinear state feedback control law

v = -* j z , - f c 2 z 2

can place the poles anywhere with proper choices of feedback gains. For example, wemay choose

v = - 2 z 2 (6.16)

resulting in the stable closed-loop dynamics

z , = - 2 z 1 + z 2

whose poles are both placed at - 2. In terms of the original state xl and x2 > t r i i s

control law corresponds to the original input

( - 2a.x2- 2 sinxj - COSJCJ sinjtj + 2xj cos Xj ) (6.17)cos {2xx)

The original state x is given from z by

xl=zl (6.18a)

jf2 = (z2-sinz1)/a (6.18b)

Since both zj and z2 converge to zero, the original state x converges to zero.

The closed-loop system under the above control law is represented in the blockdiagram in Figure 6.3. We can detect two loops in this control system, with the inner

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (229)

Sect. 6.1 Intuitive Concepts 215

loop achieving the linearization of the input-state relation, and the outer loopachieving the stabilization of the closed-loop dynamics. This is consistent with (6.14),where the control input u is seen to be composed of a nonlinearity cancellation partand a linear compensation part.

v = - k z u = u(\ , x = f(x, u)

linearization loop

pole-placement loopz = z(x)

Figure 6.3 : Input-State Linearization

A number of remarks can be made about the above control law:

• The result, though valid in a large region of the state space, is not global.The control law is not well defined when ;t[ = (JT/4 ± kn/2), k= 1,2,...Obviously, when the initial state is at such singularity points, the controllercannot bring the system to the equilibrium point.

• The input-state linearization is achieved by a combination of a statetransformation and an input transformation, with state feedback used in both.Thus, it is a linearization by feedback, or feedback linearization. This isfundamentally different from a Jacobian linearization for small rangeoperation, on which linear control is based.

• In order to implement the control law, the new state components (z(, z2)must be available. If they are not physically meaningful or cannot bemeasured directly, the original state x must be measured and used to computethem from (6.12).

• Thus, in general, we rely on the system model both for the controllerdesign and for the computation of z. If there is uncertainty in the model, e.g.,uncertainty on the parameter a, this uncertainty will cause error in thecomputation of both the new state z and of the control input u, as seen in(6.12) and (6.14).

• Tracking control can also be considered. However, the desired motion thenneeds to be expressed in terms of the full new state vector. Complex

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (230)

216 Feedback Linearization Chap. 6

computations may be needed to translate the desired motion specification (interms of physical output variables) into specifications in terms of the newstates.

With the above successful design in mind, it is interesting to extend the input-state linearization idea to general nonlinear systems. Two questions arise when onespeculates such generalizations:

• What classes of nonlinear systems can be transformed into linear systems?

• How to find the proper transformations for those which can?

These questions are systematically addressed in section 6.3.

6.1.3 Input-Output Linearization

Let us now consider a tracking control problem. Consider the system

i = f(x,K) (6.19a)

y = h(x) (6.19b)

and assume that our objective is to make the output y(t) track a desired trajectory yd(i) f

while keeping the whole state bounded, where yj{t) and its time derivatives up to a jsufficiently high order are assumed to be known and bounded. An apparent difficulty 'with this model is that the output y is only indirectly related to the input u, through the ,state variable x and the nonlinear state equations (6.19). Therefore, it is not easy to Isee how the input u can be designed to control the tracking behavior of the output y.However, inspired by the results of section 6.1.1, one might guess that the difficulty ofthe tracking control design can be reduced if we can find a direct and simple relationbetween the system output y and the control input u. Indeed, this idea constitutes theintuitive basis for the so-called input-output linearization approach to nonlinearcontrol design. Let us again use an example to demonstrate this approach.

Consider the third-order system J

i [ = sinjc2+ (JC2 + 1) x3 (6.20a) *

x2 = X[5 + x3 (6.20b) •

i 3 = xx2 + u (6.20c)

y = x{ (6.20d)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (231)

Sect. 6.1 Intuitive Concepts 217

To generate a direct relationship between the output y and the input u, let usdifferentiate the output y

y = xx = sin x2 + (x2 + 1) x3

Since y is still not directly related to the input u, let us differentiate again. We nowobtain

y = (x2+l) u +/j(x) (6.21)

where fx(x) is a function of the state defined by

fx(x) = (xx5 + x3) (x3 + cosx2) + (x2+l )xx

2 (6.22)

Clearly, (6.21) represents an explicit relationship between y and u. If we choose thecontrol input to be in the form

1 (v-/i) (6.23)x2+

where v is a new input to be determined, the nonlinearity in (6.21) is canceled, and weobtain a simple linear double-integrator relationship between the output and the newinput v,

The design of a tracking controller for this double-integrator relation is simple,because of the availability of linear control techniques. For instance, lettinge = y(t) ~ >^(0 be the tracking error, and choosing the new input v as

v=yd-kxe-k2'e (6.24)

with £j and k2 being positive constants, the tracking error of the closed loop system isgiven by

e + k2e + kle = O (6.25)

which represents an exponentially stable error dynamics. Therefore, if initiallye(0) = e(0) = 0, then e{t) = 0,V t > 0 , i.e., prefect tracking is achieved; otherwise, e(t)converges to zero exponentially.

Note that

• The control law is defined everywhere, except at the singularity points suchthat x2 = - 1 .

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (232)

218 Feedback Linearization Chap. 6

• Full state measurement is necessary in implementing the control law,because the computations of both the derivative y and the inputtransformation (6.23) require the value of x.

The above control design strategy of first generating a linear input-outputrelation and then formulating a controller based on linear control is referred to as theinput-output linearization approach, and it can be applied to many systems, as will beseen in section 6.4 for SISO systems and in section 6.5 for MIMO systems. If weneed to differentiate the output of a system r times to generate an explicit relationshipbetween the output y and input u, the system is said to have relative degree r. Thus,the system in the above example has relative degree 2. As will be shown soon, thisterminology is consistent with the notion of relative degree in linear systems (excessof poles over zeros). As we shall see later, it can also be shown formally that for anycontrollable system of order n, it will take at most n differentiations of any output forthe control input to appear, i.e., r<n. This can be understood intuitively: if it tookmore than n differentiations, the system would be of order higher than n; if the controlinput never appeared, the system would not be controllable.

At this point, one might feel that the tracking control design problem posed atthe beginning has been elegantly solved with the control law (6.23) and (6.24).However, one must remember that (6.25) only accounts for part of the closed-loopdynamics, because it has only order 2, while the whole dynamics has order 3 (thesame as that of the plant, because the controller (6.23) introduces no extra dynamics).Therefore, a part of the system dynamics (described by one state component) has beenrendered "unobservable" in the input-output linearization. This part of the dynamicswill be called the internal dynamics, because it cannot be seen from the external input-output relationship (6.21). For the above example, the internal state can be chosen tobe x3 (because x3 , y , and y constitute a new set of states), and the internal dynamics isrepresented by the equation

h = x\2 + ~l—r (Sd(O - k\e ~ k-ie +/i) (6-2 6)X *

If this internal dynamics is stable (by which we actually mean that the states remainbounded during tracking, i.e., stability in the BIBO sense), our tracking control designproblem has indeed been solved. Otherwise, the above tracking controller ispractically meaningless, because the instability of the internal dynamics would implyundesirable phenomena such as the burning-up of fuses or the violent vibration ofmechanical members. Therefore, the effectiveness of the above control design, basedon the reduced-order model (6.21), hinges upon the stability of the internal dynamics.

Let us now use some simpler examples to show that internal dynamics are

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (233)

Sect. 6.1 Intuitive Concepts 219

stable for some systems (implying that the previous design approach is applicable),and unstable for others (implying the need for a different control design).

Example 6.3: Internal dynamics

Consider the nonlinear system

+ u(6.27a)

u

y = xx (6.27b)

Assume that the control objective is to make y track yd{t). Differentiation of y simply leads to the

first state equation. Thus, choosing the control law

» = - *2 ~ e(t) + yd(t) (6.28)

yields exponential convergence of e to zero

k + e = 0 (6.29)

The same control input is also applied to the second dynamic equation, leading to the internal

dynamics

x2 + x2i=yd-e (6.30)

which is, characteristically, non-autonomous and nonlinear. However, in view of the facts that e

is guaranteed to be bounded by (6.29) and yd is assumed to be bounded, we have

\yd(t)-e\ < D

where D is a positive constant. Thus, we can conclude from (6.30) that \x2 I <D 1 / 3 (perhaps

after a transient), since x2 < 0 when x2 > D1 '3 , and x2 > 0 when x2 < — D1 '3 .

Therefore, (6.28) does represent a satisfactory tracking control law for the system (6.27),

given any trajectory yd(t) whose derivative yd(t) is bounded. D

Conversely, one can easily show (Exercise 6.2) that if the second state equationin (6.27) is replaced by Xj = — u, then the resulting internal dynamics is unstable.

Finally, let us remark that, although the input-output linearization is motivatedin the context of output tracking, it can also be applied to stabilization problems. Forexample, if y^(t) = 0 is the desired trajectory for the above system, the two states y andy of the closed-loop system will be driven to zero by the control law (6.28), implyingthe stabilization of the whole system provided that the internal dynamics is stable. In

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (234)

220 Feedback Linearization Chap. 6

addition, two useful remarks can be made about using input-output linearization forstabilization design. First, in stabilization problems, there is no reason to restrict thechoice of output v = h{x) to be a physically meaningful quantity (while in trackingproblems the choice of output is determined by the physical task). Any function of xmay be used to serve as an artificial output (a designer output) to generate a linearinput-output relation for the purpose of stabilization design. Second, different choicesof output function leads to different internal dynamics. It is possible for one choice ofoutput to yield a stable internal dynamics (or no internal dynamics) while anotherchoice of output would lead to a unstable one. Therefore, one should choose, ifpossible, the output function to be such that the associated internal dynamics is stable.

A special case occurs when the relative degree of a system is the same as itsorder, i.e., when the output y has to be differentiated n times (with n being the systemorder) to obtain a linear input-output relation. In this case, the variables y, y,...,y(n-\) m a y be usecj a s a n e w s e t of s t a j e variables for the system, and there is nointernal dynamics associated with this input-output linearization. Thus, in this case,input-output linearization leads to input-state linearization, and both state regulationand output tracking (for the particular output) can be achieved easily.

THE INTERNAL DYNAMICS OF LINEAR SYSTEMS

We must admit that it is only due to the simplicity of the system that the internaldynamics in Example 6.3 has been shown to be stable. In general, it is very difficult todirectly determine the stability of the internal dynamics because it is nonlinear, non-autonomous, and coupled to the "external" closed-loop dynamics, as seen in (6.26).Although a Lyapunov or Lyapunov-like analysis may be useful for some systems, itsgeneral applicability is limited by the difficulty of finding a Lyapunov function, asdiscussed in chapters 3 and 4. Therefore, we naturally want to seek simpler ways ofdetermining the stability of the internal dynamics. An examination of how the conceptof internal dynamics translates in the more familiar context of linear systems proveshelpful to this purpose.

Let us start by considering the internal dynamics of some simple linear systems.

Example 6.4: Internal dynamics in two linear systems

Consider the simple controllable and observable linear system

X2

u

u(6.31a)

(6.31b)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (235)

Sect. 6.1 Intuitive Concepts 221

where y(t) is required to track a desired output yd(t). With one differentiation of the output, we

simply obtain the first state equation

y = x2 + u

which explicitly contains «. Thus, the control law

u = -x2 + yd-(y-yd) (6.32)

yields the tracking error equation

(where e = y - yd) and the internal dynamics

W e see from these equations that while y(t) tends to yd(t) (and y(i) tends to yd(t)) , x2 remains

bounded, and so does u. Therefore, (6.32) is a satisfactory tracking controller for system (6.31).

Let us now consider a slightly different system:

(6.33a)X2

(6.33b)

The same control law as above yields the same tracking error dynamics, but now leads to the

internal dynamics

This implies that x2 , and accordingly u, both go to infinity as t —> ° ° . Therefore, (6.32) is not a

suitable tracking controller for system (6.33). CD

We are thus left wondering why the same tracking design method is applicableto for system (6.31) but not to system (6.33). To understand this fundamentaldifference between the two systems, let us consider their transfer functions, namely,for system (6.31),

Wyip) = £±±-P

and for system (6.33),

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (236)

222 Feedback Linearization Chap. 6

We see that the two systems have the same poles but different zeros. Specifically,system (6.31), for which the design has succeeded, has a left half-plane zero at - 1,while system (6.33), for which the design has failed, has an right half-plane zero at 1.

The above observation (the internal dynamics is stable if the plant zeros are inthe left-half plane, i.e., if the plant is "minimum-phase") can actually be shown to betrue for all linear systems, as we do now. This is not surprising because, for non-minimum phase systems, perfect tracking of arbitrary trajectories requires infinitecontrol effort, as seen in Example II.2.

To keep notations simple, let us consider a third-order linear system in state-space form

z = Az+ b« (6.34)

and having one zero (and hence two more poles than zeros), although the procedurecan be straightforwardly extended to systems with arbitrary numbers of poles andzeros. The system's input-output linearization can be facilitated if we first transform itinto the so-called companion form. To do this, we note from linear control that theinput/output behavior of this system can be expressed in the form

bo+bxpy = cT{pl-A)~xbu = -

c

(where p is the Laplace variable). Thus, if we define

x i = o—r "

(6.35)

x2 = xx

x3 = i2

the system can be equivalently represented in the companion form

00

l (T0 1

x\X2 + 0

1

(6.36a)

*

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (237)

Sect. 6.1

y = lbo o]

Intuitive Concepts 223

(6.36b)

Let us now perform input-output linearization based on this form. The firstdifferentiation of the output leads to

and the second differentiation leads to

y = bo'x2 + b\h ~ V*3H a\x2 - a2xi (6.37)

It is seen that the input u appears in the second differentiation, which means that therequired number of differentiations (the relative degree) is indeed the same as theexcess of poles over zeros (of course, since the input-output relation of y to u isindependent of the choice of state variables, it would also take two differentiations foru to appear if we used the original state-space equations (6.34)).

Thus, the control law

u = (aoxx1

—(-kle-kze + yd) (6.38)

where e = y - y^ , yields an exponentially stable tracking error

Since this is a second-order dynamics, the internal dynamics of our third-order systemcan be described by only one state equation. Specifically, we can use JCJ to completethe state vector, since one can easily show X|, y, and y are related to xx, x2, and x^through a one-to-one transformation (and thus can serve as states for the system). Wethen easily find from (6.36a) and (6.36b) that the internal dynamics is

xl = X2 = '~hox0

that is,

(6.39)

Since y is bounded (y = e + yd), we see that the stability of the internal dynamicsdepends on the location of the zero -bo/bi of the transfer function in (6.35). If the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (238)

224 Feedback Linearization Chap. 6

system is minimum phase, then the zero is in the left-half plane, which implies that theinternal dynamics (6.39) is stable, independently of the initial conditions and of themagnitudes of the desired yd,..., yj^ (where r is the relative degree).

A classical example of the effect of a right half-plane zero is the problem ofcontrolling the altitude of an aircraft using an elevator.

Example 6.5: Aircraft altitude dynamics

horizontal

Figure 6.4 : Dynamic characteristics of an aircraft

A schematic diagram of the dynamics of an aircraft (in the longitudinal plane) is shown in Figure

6.4. The sum of the lift forces applied to the aircraft wings and body is equivalent to a single lift

force Lw, applied at the "center of lift" CE . The center of lift does not necessarily coincide with

the center of mass CQ (with a positive d meaning that the center of mass is ahead of the center of

lift). The mass of the aircraft is denoted by m and its moment of inertia about CG is denoted by

/ . We assume that all angles are small enough to justify linear approximations, and that the

forward velocity of the aircraft remains essentially constant.

The aircraft is initially cruising at a constant altitude h = ho. To affect its vertical motion, the

elevator (a small surface located at the aircraft tail) is rotated by an angle E. This generates a

small aerodynamic force LE on the elevator, and thus a torque about CQ . This torque creates a

rotation of the aircraft about CQ, measured by an angle a . The lift force Lw applied to the wings

is proportional to a , i.e., Lw = Czwo.. Similarly, LE is proportional to the angle between the

horizontal and the elevator, i.e., LE = CZE(E-a) . Furthermore, various aerodynamic forces

create friction torques proportional to d, of the form ba. In summary, a simplified model of the

aircraft vertical motion can be written

(CZEl+Czwd)a = CZEl E (6.40a)

m'h = {CZE^Cm)a-CZEE (6.40b)

where the first equation represents the balance of moments and the second the balance of forces. 1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (239)

Sect. 6.1 Intuitive Concepts 225

Remark that the open-loop stability of the first equation, which defines the dynamics of the

angle a, depends on the sign of the coefficient (CZgl + C^d). In particular, the equation is

open-loop stable if d > 0 , i.e., if the center of mass is ahead of the center Of lift (this allows us to

understand the shape of a jumbo jet, or the fact that on small aircraft passengers are first seated on

the front rows).

To simplify notations, let now

y = 1 m=l b = 4 CZE=l Cm = 5 1 = 3 d = 0.2.

The transfer functions describing the system can then be written

§4 TE(p) p2 + 4p + 4 (p + 2)

h(p) _ \4-4p-p2 _ (6.24+/>)(2.24-p)EYTA TT^y—~, 77 T~, XTo vo.tio;

where p is the Laplace variable.

At time t = 0, a unit step input in elevator angle E is applied. From the initial and final value

theorems of linear control (or directly from the equations of motion), one can easily show that the

corresponding initial and final vertical accelerations are

M7=0+) = - 1 < 0 h(t= + °°) = 3.5 > 0

The complete time responses in h and h are sketched in Figure 6.5. We see that the aircraft starts

in the wrong direction, then recovers. Such behavior is typical of systems with a right half-plane

zero. It can be easily understood physically, and is a direct reflection of the aircraft design itself.

The initial effect of the unit step in E is to create an instantaneous downward force on the

elevator, thus creating an initial downward acceleration of the aircraft's center of mass. The unit

step in elevator angle also creates a torque about CQ , which builds up the angle a and thus

creates an increasing upward lift force on the wing and body. This lift force eventually takes over

the downward force on the elevator. Of course, such non-minimum phase behavior is important

for the pilot to know, especially when flying at low altitudes.

Let us determine the associated internal dynamics. Defining the state as x = [ a a h h]T,

the equations of motion can be written

x, = x2 (6.42a)

x2 = — 4^2 — 4 x j + 3 £ (6.42b)

x3 = x4 (6.42c)

kA = 6x, - E (6.42d)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (240)

226 Feedback Linearization Chap. 6

Im

Re

h

3.5

-1time time

(a) (b) (c)

Figure 6.5 : Pole-zero plot and step responses in h and h

The output of interest is the aircraft's altitude

y=xi

Differentiating y until the input E appears yields

y = * 3 = xA = 6xl - E

consistent with the fact that the transfer function (6.41b) has relative degree 2. Choose now a

pole-placement control law for E

where y = y - yd . Then

y + y + y = o

The corresponding internal dynamics is

x2 = -4JT2-4-Kj + 3(6jf|

that is

(6.43)

+ y)

a + 4d- 14a = 'i{-ydJry + y) (6.44)

and therefore, is unstable. Specifically, the poles on the left-hand side of (6.44) are exactly the

zeros of the transfer function (6.41b).

THE ZERO-DYNAMICS•

Since for linear systems the stability of the internal dynamics is simply determined bythe locations of the zeros, it is interesting to see whether this relation can be extendedto nonlinear systems. To do so requires first to extend the concept of zeros tononlinear systems, and then to determine the relation of the internal dynamics stability

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (241)

Sect. 6.1 Intuitive Concepts 227

to this extended concept of zeros.

Extending the notion of zeros to nonlinear systems is not a trivial proposition.Transfer functions, on which linear system zeros are based, cannot be defined fornonlinear systems. Furthermore, zeros are intrinsic properties of a linear plant, whilefor nonlinear systems the stability of the internal dynamics may depend on the specificcontrol input.

A way to approach these difficulties is to define a so-called zero-dynamics for anonlinear system. The zero-dynamics is defined to be the internal dynamics of thesystem when the system output is kept at zero by the input. For instance, for thesystem (6.27), the zero-dynamics is (from (6.30))

x 2+x 23 = 0 (6.45)

Noticing that the specification of maintaining the system output at zero uniquelydefines the required input (namely, here, u has to equal —x^ in order to keep Xjalways equal to zero), we see that the zero-dynamics is an intrinsic property of anonlinear system. The zero-dynamics (6.45) is easily seen to be asymptotically stable(by using the Lyapunov function K = x 2

2 ) .

Similarly, for the linear system (6.34), the zero-dynamics is (from (6.39))

xj +{bolbx)xl = 0

Thus, in this linear system, the poles of the zero-dynamics are exactly the zeros of thesystem. This result is general for linear systems, and therefore, in linear systems,having all zeros in the left-half complex plane guarantees the global asymptoticstability of the zero-dynamics.

The reason for defining and studying the zero-dynamics is that we want to finda simpler way of determining the stability of the internal dynamics. For linearsystems, the stability of the zero-dynamics implies the global stability of the internaldynamics: the left-hand side of (6.39) completely determines its stabilitycharacteristics, given that the right-hand side tends to zero or is bounded. In nonlinearsystems, however, the relation is not so clear. Section 6.4 investigates this question insome detail. For stabilization problems, it can be shown that local asymptotic stabilityof the zero-dynamics is enough to guarantee the local asymptotic stability of theinternal dynamics. Extensions can be drawn to the tracking problem. However, unlikethe linear case, no results on the global stability or even large range stability can bedrawn for internal dynamics of nonlinear systems, i.e., only local stability isguaranteed for the internal dynamics even if the zero-dynamics is globallyexponentially stable.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (242)

228 Feedback Linearization Chap. 6

Similarly to the linear case, we will call a nonlinear system whose zero-dynamics is asymptotically stable an asymptotically minimum phase system. Theconcept of an exponentially minimum phase system can be defined in the same way.

Two useful remarks can be made about the zero-dynamics of nonlinear systems.First, the zero-dynamics is an intrinsic feature of a nonlinear system, which does notdepend on the choice of control law or the desired trajectories. Second, examining thestability of zero-dynamics is much easier than examining the stability of internaldynamics, because the zero-dynamics only involves the internal states (while theinternal dynamics is coupled to the external dynamics and desired trajectories, as seenin (6.26)).

Example 6.6: Aircraft zero-dynamics

Given (6.44), the zero-dynamics of the aircraft of Example 6.5 is

a + 4 o - 14a = 0 (6.46)

This dynamics is unstable, confirming that the system is non-minimum phase. The poles of the

zero-dynamics are exactly the zeros of the transfer function (6.41 b).

Now model (6.40) is actually the linearization of a more general nonlinear model, applicable

at larger angles and angular rates. Since the zero-dynamics corresponding to the linearized model

simply is the linearization of the zero-dynamics corresponding to the nonlinear model, thus the

nonlinear system is also non-minimum phase, from Lyapunov's linearization method. D

To summarize, control design based on input-output linearization can be madein three steps:

• differentiate the output y until the input u appears

• choose u to cancel the nonlinearities and guarantee tracking convergence

• study the stability of the internal dynamics

If the relative degree associated with the input-output linearization is the same as theorder of the system, the nonlinear system is fully linearized and this procedure indeedleads to a satisfactory controller (assuming that the model is accurate). If the relativedegree is smaller than the system order, then the nonlinear system is only partlylinearized, and whether the controller can indeed be applied depends on the stability ofthe internal dynamics. The study of the internal dynamics stability can be simplifiedlocally by studying that of the zero-dynamics instead. If the zero-dynamics isunstable, different control strategies should be sought, only simplified by the fact thatthe transformed dynamics is partly linear.

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (243)

Sect. 6.2 Mathematical Tools 229

6.2 Mathematical Tools

The objective of the rest of this chapter is to formalize and generalize the previousintuitive concepts for a broad class of nonlinear systems. To this effect, we firstintroduce some mathematical tools from differential geometry and topology. To limitthe conceptual and notational complexity, we discuss these tools directly in thecontext of nonlinear dynamic systems (instead of general topological spaces). Notethat this section and the remainder of this chapter represent by far the mostmathematically involved part of the book. Hurried practitioners may skip it in a. firstreading, and go directly to chapters 7-9.

In describing these mathematical tools, we shall call a vector functionf: R" —> R" a vector field in R", to be consistent with the terminology used indifferential geometry. The intuitive reason for this term is that to every vector functionf corresponds a field of vectors in an ^-dimensional space (one can think of a vectorf(x) emanating from every point x). In the following, we shall only be interested insmooth vector fields. By smoothness of a vector field, we mean that the function f(x)has continuous partial derivatives of any required order.

Given a smooth scalar function h(x) of the state x, the gradient of h is denotedby Vh

The gradient is represented by a row-vector of elements (Vh): = dh I dx:. Similarly,given a vector field f(x), the Jacobian of f is denoted by Vf

vf = i?

dx

It is represented by an n x n matrix of elements (Vf)(- • = 3/j- I dx:.

LIE DERIVATIVES AND LIE BRACKETS

Given a scalar function h(x) and a vector field f(x), we define a new scalar functionLfh, called the Lie derivative (or simply, the derivative) of h with respect to f.

Definition 6.1 Let h : R" —> R be a smooth scalar function, and f: R" —> R" be asmooth vector field on R", then the Lie derivative of h with respect to f is a scalarfunction defined by Lfh = Vh f.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (244)

230 Feedback Linearization Chap. 6

Thus, the Lie derivative Lfh is simply the directional derivative of h along the kdirection of the vector f.

Repeated Lie derivatives can be defined recursively

Lfh = h

L{'h = Lf(Lf '-' h)= V(L fM h) f for i= 1,2,

Similarly, if g is another vector field, then the scalar function Lg Lf h(x) is

g

One can easily see the relevance of Lie derivatives to dynamic systems byconsidering the following single-output system

x = f(x)

y = h(x)

The derivatives of the output are

d h • . ,y = — x = Lfh

y =dx

and so on. Similarly, if V is a Lyapunov function candidate for the system, itsderivative V can be written as Lf V.

Let us move on to another important mathematical operator on vector fields, theLie bracket.

Definition 6.2 Let f and g be two vector fields on R". The Lie bracket off and g is athird vector field defined by

[f, g] = Vg f - Vf g

The Lie bracket [f, g] is commonly written as adf g (where ad stands for "adjoint").Repeated Lie brackets can then be defined recursively by

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (245)

Sect. 6.2 Mathematical Tools 231

adf g = g

adj g = [t, adf1'1 g] for / = 1 , 2 ,

Example 6.7: The system (6.11) can be written in the form

with the two vector fields f and g defined by

f =

Their Lie bracket can be computed as

[f,g]=:

g(x) =COS(2A:|)

-2sin(2x[)

cosxjcos(2xj

) -

— 2xj + ax2 + si

—X2C0SXJ

acos(2xj)

2sin(2x,)(-2x,

nx t —2 + cosxj

x2 sin Xj

+ ax2 + sinx()

a

— COSXj

cos(2x[)

r

The following lemma on Lie bracket manipulation will be useful later.

Lemma 6.1 Lie brackets have the following properties

(i) bilinearity:

[f, c^g, + a2g2] = ax [f, gj] + a2[f, g2]

where f, f j , f2 , g , g] and g2 are smooth vector fields, and CC[ and a 2 areconstant scalars.

(it) skew-commutativity:

[f, g] = - [g, f]

(Hi) Jacobi identity:

Lad(gh = L{Lgh-LgLfh

where h(\) is a smooth scalar function of x.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (246)

232 Feedback Linearization Chap. 6

Proof. The proofs of the first two properties are straightforward (Exercise 6.6). Let us derive the

third property, which can be rewritten as

The left-hand side of the above equation can be expanded as

while the right-hand side can be expanded as

where 32/i / 3x2 is the Hessian of h, a symmetric matrix. D

The Jacobi identity can be used recursively to obtain useful technical identities.Using it twice yields

Lad(2g h ~ Ladf(adfg) h ~ LfLad(g

h ~

(6.47)

- LgLfh] - [LfLg -

Lg h - 2Lf Lg Lf h Lf2 h

Similar identities can be obtained for higher-order Lie brackets.

DIFFEOMORPHISMS AND STATE TRANSFORMATIONS

The concept of diffeomorphism can be viewed as a generalization of the familiarconcept of coordinate transformation. It is formally defined as follows:

Definition 6.3 A function <|>: R" —> R", defined in a region Q, is called adiffeomorphism if it is smooth, and if its inverse §~l exists and is smooth.

If the region Q is the whole space R", then <))(x) is called a globaldiffeomorphism. Global diffeomorphisms are rare, and therefore one often looks forlocal diffeomorphisms, i.e., for transformations defined only in a finite neighborhoodof a given point. Given a nonlinear function <(>(x), it is easy to check whether it is alocal diffeomorphism by using the following lemma, which is a straightforwardconsequence of the well-known implicit function theorem.

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (247)

Sect. 6.2 Mathematical Tools 233

Lemma 6.2 Let (j>(xj be a smooth function defined in a region £1 in Rn. // theJacobian matrix V<(» is non-singular at a point \ = x0 of Q., then §(x) defines a localdiffeomorphism in a subregion ofCi.

A diffeomorphism can be used to transform a nonlinear system into anothernonlinear system in terms of a new set of states, similarly to what is commonly donein the analysis of linear systems. Consider the dynamic system described by

y = h(\)

and let a new set of states be defined by

z = (|>(x)

Differentiation of z yields

z ! * ! <«x)6 x ox

One can easily write the new state-space representation as

i = f*(z) + g*(z)u

y = h*(z)

where x = (^'(z) has been used, and the functions f *, g* and h* are defined obviously.

Example 6.8; A non-global diffeomorphism

The nonlinear vector function

= <Kx) =3sinx-,

(6.48)

is well defined for all x t and x2. Its /acobian matrix is

9x 0

10x,x2

3cosx2

which has rank 2 at x = (0, 0). Therefore, Lemma 6.2 indicates that the function (6.48) defines a

local diffeomorphism around the origin. In fact, the diffeomorphism is valid in the region

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (248)

234 Feedback Linearization Chap. 6

Q = ( ( x , , x 2 ) , |x2 |<rc/2)

because the inverse exists and is smooth for x in this region. However, outside this region, § does

not define a diffeomorphism, because the inverse does not uniquely exist. Q

THE FROBENIUS THEOREM

The Frobenius theorem is an important tool in the formal treatment of feedbacklinearization for «tn-order nonlinear systems. It provides a necessary and sufficientcondition for the solvability of a special class of partial differential equations. Beforepresenting the precise statement of the theorem, let us first gain a basic understandingby discussing the case n = 3.

Consider the set of first-order partial differential equations

djtdx. 8l +

>X2

dhdx,

dh3*3

(6.49a)

(6.49b)

where and C| ,^2,^3) (/= 1,2,3) are known scalar functions ofxl>x2>x3> a n d h{x\,X2,x$ is an unknown function. Clearly, this set of partialdifferential equations is uniquely defined by the two vectorsf = [/, f2 / 3 ]T , g = [ gx g2 g3 ]T. If a solution h(xx ,x2,x3) exists for the above

partial differential equations, we shall say the set of vector fields {f , g} is completelyintegrable.

The question now is to determine when these equations are solvable. This is notobvious at all, a priori. The Frobenius theorem provides a relatively simple condition:Equation (6.49) has a solution h(x\ ,^2,^3) if, and only if, there exists scalar functions

3) s u c n m a t

[f, g] = « ! f + a 2 g

i.e., if the Lie bracket of f and g can be expressed as a linear combination of f and g.This condition is called the involutivity condition on the vector fields ( f , g | •Geometrically it means that the vector [f, g] is in the plane formed by the two vectors fand g. Thus, the Frobenius theorem states that the set of vector fields { f , g | iscompletely integrable if, and only if, it is involutive. Note that the involutivitycondition can be relatively easily checked, and therefore, the solvability of (6.49) canbe determined accordingly.

Let us now discuss the Frobenius theorem in the general case, after giving

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (249)

Sect. 6.2 Mathematical Tools 235

formal definitions of complete integrability and involutivity.

Definition 6.4 A linearly independent set of vector fields {fj, f2 , . . . , fm} on R" issaid to be completely integrable if, and only if, there exist n—m scalar functionsh^ix), h2(\),...., hn_m{x) satisfying the system of partial differential equations

Vh; fj = 0 (6.50)

where 1 < / < n—m , 1 <j < m, and the gradients Vhj are linearly independent.

Note that with the number of vectors being m and the dimension of the associatedspace being n, the number of unknown scalar functions hv involved is (n-m) and thenumber of partial differential equations is m(n-m).

Definition 6.5 A linearly independent set of vector fields {fj , f2,..., fm ) is said tobe involutive if, and only if, there are scalar functions a^ : R" —> R such that

[f,-, f.](x) = $ W x ) f*(x) V/,7 (6.51)

Involutivity means that if one forms the Lie bracket of any pairs of vector fields fromthe set {fj , f2 , . . . , fm ), then the resulting vector field can be expressed as a linearcombination of the original set of vector fields. Note that

• Constant vector fields are always involutive. Indeed, the Lie bracket oftwo constant vectors is simply the zero vector, which can be triviallyexpressed as linear combination of the vector fields.

• A set composed of a single vector f is involutive. Indeed,

[f,f] = (Vf)f-(Vf)f=O

• From Definition 6.5, checking whether a set of vector fields {fj ,...., fm}is involutive amounts to checking whether

rank(f,(x) .... fm(x)) = rank(f,(x) ... fM(x) [f,-, fy](x))

for all x and all i,j.

We can now state the Frobenius theorem formally.

Theorem 6.1 (Frobenius) Let f [ , f2 , . . . , fm be a set of linearly independent vectorfields. The set is completely integrable if, and only if, it is involutive.

Example 6.9: Consider the set of partial differential equations

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (250)

236 Feedback Linearization Chap. 6

4,3!*-!* =o3 Xj 3 x2

3 h , 9 O . 3 / i - 3ft „- * , — + (X 32_3X 2) + 2 x 3 - — = 0

O Xy OX2 0JC3

The associated vector fields are jf, , f2) with

f, = [4*3 - 1 O f f2 = [ - JC , ( x 32 - 3 x 2 ) lx{f

In order to determine whether this set of partial differential equations is solvable (or whether

(fj f2! is completely integrable), let us check the involutivity of the set of vector fields [fl f2).

One easily finds that

[fl,f2] = [-12x3 3 Of

Since [fj, f2] = — 3 fj + 0f2 , this set of vector fields is involutive. Therefore, the two partial

differential equations are solvable. D

6.3 Input-State Linearization of SISO Systems

In this section, we discuss input-state linearization for single-input nonlinear systemsrepresented by the state equations

i = f(x) + g(x)K (6.52)

with f and g being smooth vector fields. We study when such systems can belinearized by state and input transformations, how to find such transformations, andhow to design controllers based on such feedback linearizations.

Note that systems in the form (6.52) are said to be linear in control or affine. Itis useful to point out that if a nonlinear system has the form

x = f(x) + g(x) w[u + <|>(x)]

with w being an invertible scalar function and (|> being an arbitrary functional, a simplevariable substitution v = w[u + <|>(x)] puts the dynamics into the form of (6.52). Onecan design a control law for v and then compute u by inverting w, i.e.,

DEFINITION OF INPUT-STATE LINEARIZATION

In order to proceed with a detailed study of input-state linearization, a formaldefinition of this concept is necessary:

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (251)

Sect. 6.3 Input-State Linearization o/SISO Systems 237

Definition 6.6 A single-input nonlinear system in the form (6.52), with f(x) and g(x)being smooth vector fields on W, is said to be input-state linearizable if there exists aregion £2 in R", a diffeomorphism <|> : O —-> R", and a nonlinear feedback control law

u = a(x) + p(x)v (6.53)

such that the new state variables z = <{t(x) and the new input v satisfy a linear time-invariant relation

where

z

A

= Az+ bv

=

"00

1

0 .

1 .

0 .

0 .

. 0

. 1

. 0

(6.54)

b =

The new state z is called the linearizing state, and the control law (6.53) is called thelinearizing control law. To simplify notations, we will often use z to denote not onlythe transformed state, but the diffeomorphism <|> itself, i.e. write

z = z(x)

This slight abuse of notations should not create any confusion.

Note that the transformed linear dynamics has its A matrix and b vector of aspecial form, corresponding to a linear companion form. However, generality is notlost by restricting ourselves to this special linear equivalent dynamics, because anyrepresentation of a linear controllable system is equivalent to the companion form(6.54) through a linear state transformation and pole placement. Therefore, if (6.53)can be transformed into a linear system, it can be transformed into the form prescribedby (6.54) by using additional linear transformations in state and input.

We easily see from the canonical form (6.54) that feedback linearization is aspecial case of input-output linearization, where the output function leads to a relativedegree n. This means that if a system is input-output linearizable with relative degreen, it must be input-state linearizable. On the other hand, if a system is input-statelinearizable, with the first new state Z\ representing the output, the system is input-output linearizable with relative degree n. Therefore, we can summarize the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (252)

238 Feedback Linearization Chap. 6

relationship between input-output linearization and input-state linearization as follows:

Lemma 6.3 An nth-order nonlinear system is input-state linearizable if, and only if,there exists a scalar function Zj(x) such that the system's input-output linearizationwith Z[(x) as output function has relative degree n.

Note, however, that the above lemma provides no guidance about how to find thedesirable output function Z[(x).

CONDITIONS FOR INPUT-STATE LINEARIZATION

At this point, a natural question is: can all nonlinear state equations in the form of(6.52) be input-state linearized? If not, when do such linearizations exist? Thefollowing theorem provides a definitive answer to that question, and constitutes one ofthe most fundamental results of feedback linearization theory.

Theorem 6.2 The nonlinear system (6.52), with ffxj and g(x) being smooth vectorfields, is input-state linearizable if, and only if, there exists a region Q such that thefollowing conditions hold:

• the vector fields {g, adf g ,...., adfn~^ g} are linearly independent in Q

• the set {g, adf g ,.. . , ad{"~^ g} is involutive in Q

Before proving this result, let us make a few remarks about the above conditions:

• The first condition can be interpreted as simply representing acontrollability condition for the nonlinear system (6.52). For linear systems,the vector fields (g, adfg,...., adf

n~l g } become { b , Ab ,...., A""1 b ) ,and therefore their independence is equivalent to the invertibility of thefamiliar linear controllability matrix. It is also easy to show that if a system'slinear approximations in a closed connected region Q in R" are allcontrollable, then, under some mild smoothness assumptions, the system canbe driven from any point in Q to any point in O. However, as mentioned inchapter 3, a nonlinear system can be controllable while its linearapproximation is not. The first condition above can be shown to represent ageneralized controllability condition which also accounts for such cases.

• The involutivity condition is less intuitive. It is trivially satisfied for linearsystems (which have constant vector fields), but not generically satisfied inthe nonlinear case.

Let us now prove the above theorem. We first state a technical lemma.

j

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (253)

Sect. 6.3 Input-State Linearization ofSISO Systems 239

Lemma 6.4 Let z(x) be a smooth function in a region il. Then, in Q, the set of equations

Lgz = LgLfz = ... = LgL{kz = O (6.55a)

is equivalent to

Lgi = Ladfgz = ... = Ladfkgz = O (6.55b)

for any positive integer k.

Proof: Let us show that (6.55a) implies (6.55b).

When k = 0, the result is obvious. When k = 1, we have from Jacobi's identity (Lemma 6.1)

Lad,%z = LfLgz-LgLfz = 0-0 = 0

When k = 2, we further have, using Jacobi's identity twice as in (6.47)

Ladi%z = Lf2Lgz-2LtLgLfz + LgLt

2z = 0 -0 + 0 = 0

Repeating this procedure, we can show by induction that (6.55a) implies (6.55b) for any k.

One proceeds similarly to show that (6.55b) implies (6.55a) (by using Jacobi's identity the

other way around). D

We are now ready for the proof of Theorem 6.2 itself.

Proof of Theorem 6.2: Let us first prove the necessity of the conditions. Assume that there exist

a state transformation z = z(x) and an input transformation u = a(x) + P(x) v such that z and v

satisfy (6.54). Expanding the first line of (6.54), we obtain

dz,z

Proceeding similarly with the other components of z leads to a set of partial differential equations

3 z, d z,—if+ - i g « = z23x 3x

d Zj d z2- — f + - — gi/ = z39x dx

dzne dnf + g W = V

d X dx

Since Z| ,... , zn are independent of u, while v is not, we conclude from the above equations that

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (254)

240 Feedback Linearization

' = 1.2,..., «

Chap. 6

(6.56a) >l

(6.56b) | |

IThe above equations on the z,- can be compressed into a set of constraint equations on 2,alone. Indeed, using Lemma 6.4, equation (6.56a) implies that

Vzlad(kg = 0 k = 0, 1,2,... ,n-2

Furthermore by proceeding as in the proof of Lemma 6.4, we can show

(6.57a)

This implies that

Vzlad("-lg * 0 (6.57b)

The first property we can now infer from (6.57) is that the vector fieldsg, adf g ,... , adf"~\ g must be linearly independent. Indeed, if for some number i (i < n-1), thereexisted scalar functions <X|(x),..., c t^x) such that

/-I

we would have

n-2adf" ' 8 = o.kad{

k g

This, together with (6.57a), would imply that

n-2

k=n-i-\

a contradiction to (6.57b).

The second property we can infer from (6.57) is that the vector fields are involutive. Thisfollows from the existence of a scalar function z{ satisfying the n- 1 partial differential equationsin (6.57a), and from the necessity part of the Frobenius theorem. Thus, we have completed thenecessity part of the proof of Theorem 6.2.

Let us now prove that the two conditions in Theorem 6.2 are also sufficient for the input-statelinearizability of the nonlinear system in (6.52), i.e., that we can find a state transformation and aninput transformation such that (6.54) is satisfied. The reasoning is as follows. Since theinvolutivity condition is satisfied, from Frobenius theorem there exists a non-zero scalar functionz,(x) satisfying

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (255)

Sect. 6.3 Input-State Linearization ofSISO Systems 241

LS z\ = Lodfg H = - = Lodt"-\ z\ = ° (6-58)

From Lemma 6.4, the above equations can be written

Lg z, = Lg Ltzx = ... = Lg L{"-2 z, = 0 (6.59)

This means that if we use z = [ z{ Lf Zj ... L{"~' zj ]T as a new set of state variables, the first

n— 1 state equations verify

zk = zk+\ * = l , . . . . n - l

while the last state equation is

zn = Lfzx+L%Lf~xzxu (6.60)

Now the question is whether L Lt"~izl can be equal to zero. Since the vector fields

{ g, adfg,..., adf"~lg \ are linearly independent in £1, and noticing, as in the proof of Lemma

6.4, that (6.58) also leads to

we must have

W-'gM*) * 0 VxeQ (6.61)

Otherwise, the non-zero vector Vz, would satisfy

V z l t g adi% - ^ f " " ' g ] = 0

and thus would be orthogonal to n linearly independent vectors, a contradiction.

Therefore, by taking the control law to be

u = (~L("zl + v)/(L%Lt"-izl)

equation (6.60) simply becomes

'zn = v

which shows that the input-output linearization of the nonlinear system has been achieved. d

HOW TO PERFORM INPUT-STATE LINEARIZATION

Based on the previous discussion, the input-state linearization of a nonlinear system

can be performed through the following steps:

• Construct the vector fields g, adf g,...., adfn~ 'g for the given system

• Check whether the controllability and involutivity conditions are satisfied

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (256)

242 Feedback Linearization Chap. 6

• If both are satisfied, find the first state zy (the output function leading toinput-output linearization of relative degree n) from equations (6.58), i.e.,

= 0 ; = 0,... , n - 2 (6.62a)

(6.62b)

• Compute the state transformation z(x) = [ zj Lfz\ ... Lfn *z y }T and the

input transformation (6.53), with

a(x) = - (6.63a)

(6.63b)

Let us now demonstrate the above procedure on a simple physical example[Marino and Spong, 1986; Spong and Vidyasagar, 1989].

Example 6.10: Consider the control of the mechanism in Figure 6.6, which represents a link

driven by a motor through a torsional spring (a single-link flexible-joint robot), in the vertical

plane. Its equations of motion can be easily derived as

k(ql-q2) =

J'q2-k(ql-q2) = u

(6.64a)

(6.64b)

Figure 6.6 : A flexible-joint mechanism

Because nonlinearities (due to gravitational torques) appear in the first equation, while the control

input u enters only in the second equation, there is no obvious way of designing a large range

controller. Let us now consider whether input-state linearization is possible.

First, let us put the system's dynamics in a state-space representation. Choosing the state

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (257)

Sect. 6.3

vector as

Input-State Linearization ofSISO Systems 243

x = U l <?1 <72 <?2i

the corresponding vector fields f and g can be written

kf = [x2 sinx, - - ( X | - J

g = [0 o o i ] r

Second, let us check the controllability and involutivity conditions. The controllabilitymatrix is obtained by simple computation

0 0 0 - - 1

[g ad(% ad2fg ad3fg] =

77

It has rank 4 for k>0,IJ <°°. Furthermore, since the vector fields ) g, adf g, ad{ 2g) are

constant, they form an involutive set. Therefore, the system in (6.64) is input-state linearizable.

Third, let us find out the state transformation z = z(x) and the input transformationu = <x(x) + j5(x) v so that input-state linearization is achieved. From (6.62), and given the above

expression of the controllability matrix, the first component Z[ of the new state vector z shouldsatisfy

3 x-,= 0

d x-.= 0

d xA

= 0

Thus, zj must be a function of jrf only. The simplest solution to the above equations is

z, = JC.

The other states can be obtained from z.

z3 = Vz2 f = - —i— smxx - - (x, - x3)

(6.65a)

(6.65b)

(6.65c)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (258)

244 Feedback Linearization

z4 = Vz3 f = - (x2 cosx, - j (x2 - x4)

Accordingly, the input transformation is

u = (v-Vz4f)/(Vz4g)

which can be written explicitly as

Chap. 6

(6.65d)

(6.66)

where

a(\) = —f—sm/

* * MgL( - + - + —f—+I J

As a result of the above state and input transformations, we end up with the following set of linear

equations

24=V

thus completing the input-state linearization.

Finally, note that

• The above input-state linearization is actually global, because the diffeomorphism z(x)

and the input transformation are well defined everywhere. Specifically, the inverse of the

state transformation (6.65) is

MgL

which is well defined and differentiable everywhere. The input transformation (6.66) is

also well defined everywhere, of course.

• In this particular example, the transformed variables have physical meanings. We see that

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (259)

Sect. 6.3 Input-State Linearization of SISO Systems 245

Z] is the link position, z2 the link velocity, z3 the link acceleration, and z4 the link jerk. This

further illustrates our earlier remark that the complexity of a nonlinear physical model is

strongly dependent on the choice of state variables.

• In hindsight, of course, we also see that the same result could have been derived simply

by differentiating equation (6.64a) twice, i.e., from the input-output linearization

perspective of Lemma 6.3. D

Note that inequality (6.62b) can be replaced by the normalization equation

VZlad("-lg = 1

without affecting the input-state linearization. This equation and (6.62a) constitute atotal of n linear equations,

[adfg ad\g ...

dzl/dxn

' 0 '0

01

Given the independence condition on the vector fields, the partial derivativesdz^/dxy , , dzy/dx,, can be computed uniquely from the above equations. The statevariable Zj can then be found, in principle, by sequentially integrating these partialderivatives. Note that analytically solving this set of partial differential equations forzj may be a nontrivial step (although numerical solutions may be relatively easy dueto the recursive nature of the equations).

CONTROLLER DESIGN BASED ON INPUT-STATE LINEARIZATION

With the state equation transformed into a linear form, one can easily designcontrollers for either stabilization or tracking purposes. A stabilization example hasalready been provided in the intuitive section 6.1, where v is designed to place thepoles of the equivalent linear dynamics, and the physical input u is then computedusing the corresponding input transformation. One can also design trackingcontrollers based on the equivalent linear system, provided that the desired trajectorycan be expressed in terms of the first linearizing state component zj.

Consider again the flexible link example. Its equivalent linear dynamics can beexpressed as

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (260)

246 Feedback Linearization Chap. 6

Assume that it is desired to have the link position Zj track a prespecified trajectory). The control law

a3 ~ ao

leads to the tracking error dynamics(where

The above dynamics is exponentially stable if the positive constants a(- are chosenproperly. To find the physical input u, one then simply uses (6.66).

6.4 Input-Output Linearization of SISO Systems

In this section, we discuss input-output linearization of single-input nonlinear systemsdescribed by the state space representation

V = h(\)

(6.67a)

(6.67b)

where y is the system output. By input-output linearization we mean the generation ofa linear differential relation between the output y and a new input v (v here is similarto the equivalent input v in input-state linearization). Specifically, we shall discuss thefollowing issues:

• How to generate a linear input-output relation for a nonlinear system?

• What are the internal dynamics and zero-dynamics associated with theinput-output linearization?

• How to design stable controllers based on input-output linearizations?

GENERATING A LINEAR INPUT-OUTPUT RELATION

As discussed in section 6.1.3, the basic approach of input-output linearization issimply to differentiate the output function y repeatedly until the input u appears, andthen design u to cancel the nonlinearity. However, in some cases, the second part ofthe approach may not be carried out, because the system's relative degree isundefined.

The Case of Well Defined Relative Degree

Let us place ourselves in a region (an open connected set) £2X in the state space. Using

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (261)

Sect. 6.4 Input-Output Linearization of SISO Systems 247

the notations of differential geometry, the process of repeated differentiation meansthat we start with

y = Vh (f + g u) = Lf h(x) + Lg h(x) u

If Lg h(x) * 0 for some x = x0 in Q x , then, by continuity, that relation isverified in a finite neighborhood £2 of xg . In D, the input transformation

u = —!— (-Lfh + v)Lgh

results in a linear relation between y and v, namely y = v.

If Lg h(x) = 0 for all x in £2X , we can differentiate y to obtain

If Lg Lf /z(x) is again zero for a// x in Qx , we shall differentiate again and again,

jW = L f ' h(x) + Lg L f ' - * /;(x)M

until for some integer r

LgL{r~l h(x)*0

for some x = x0 in ilx . Then, by continuity, the above relation is verified in a finiteneighborhood Q of xo . In Q, the control law

u = ! (-Lfrh + v) (6.68)

LgLf'""1/!

applied to

>>('•> = L f ' h(x) + Lg L f ' - ! /i(x) M (6.69)

yields the simple linear relation

3-W = v (6.70)

As discussed in section 6.1.3, the number r of differentiations required for theinput u to appear is called the relative degree of the system, an extension of the usualdefinition of relative degree for linear systems. As also noticed earlier, and as weshall soon formalize, one necessarily has r < n (where n is the system order). Ifr = n , then input-output linearization actually yields input-state linearization in O., asstated in Lemma 6.3.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (262)

248 Feedback Linearization Chap. 6

Based on the above procedure, we can give the following formal definition.

Definition 6.7 The SISO system is said to have relative degree r in a region O if,V X G Q

LgLf'/j(x) = 0 0<i<r-l (6.71a)

LgLf'-'/iCx) * 0 (6.71b)

The Case of Undefined Relative Degree

It is often the case that one is a priori interested in the properties of the system about aspecific operating point x0. The definition of the relative degree then requiresparticular care.

As before, let us differentiate the output y until the input u appears (i.e., until thecoefficient of u is not identically zero in a neighborhood of x0). Then, if thecoefficient of u is nonzero at x0

this also implies, by continuity, that (6.71b) is verified over a finite neighborhood ofx0 . We shall then say that the system has relative degree r at the point x0 . |

However, it is possible that when the input appears, its coefficient •Lg Lfr~' h(x) is zero at xo , but nonzero at some points x arbitrarily close to x0 . |The relative degree of the nonlinear system is then undefined at x 0 . This case isillustrated by the following simple example.

Example 6.1 IT Consider the system

x = p (x, x) + u

where p is some smooth nonlinear function of the state x = [ x x ]T. If

y = x

is defined as the output of interest, then the system is obviously in companion form, with relative

degree 2.

However, if instead one takes

as output, then

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (263)

Sect. 6.4 Input-Output Linearization ofSISO Systems 249

y = 2 x'x

y = 2 / J + 2x2 = 2jt(p+ «) + 2x2

In other words,

LgL{h = 2x

and therefore, at x = 0 and for this choice of output, the system has neither relative degree 1 nor

relative degree 2. E3

In some particular cases, as above, a simple change of output will allow one todefine an equivalent but easily solvable control problem. In general, however, input-output linearization at a point cannot be straightforwardly achieved when the relativedegree is undefined. In the rest of this section, we consider only systems having awell defined relative degree in a region Q (which shall often be a neighborhood of anoperating point of interest x0 ).

NORMAL FORMS

When the relative degree r is defined and r < n, the nonlinear system (6.67) can betransformed, using y,y,..., y(r~V as part of the new state components, into a so-called "normal form", which shall allow us to take a more formal look at the notions ofinternal dynamics and zero-dynamics introduced in section 6.1.3. Let

|i = [n, H2 ... \ir]T = [y y ... y(.r-»f (6.72)

In a neighborhood Q of a point xo , the normal form of the system can be written as

V- =

a(\i,

with the output defined as

(6.73a)

(6.73b)

(6.74)

The (i(- and i|/- are referred to as normal coordinates or normal states in Q (or at x0 ).Note that the companion form subsystem (6.73a) is simply another expression of(6.69), while the subsystem (6.73b) does not contain the system input u.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (264)

250 Feedback Linearization Chap. 6

To show that the nonlinear system (6.67) can indeed be transformed into thenormal form (6.73), we have to show not only that such a coordinate transformationexists, but also that it is a true state transformation. In other words, we need to showthat we can construct a (local) diffeomorphism

such that (6.73) is verified. According to Lemma 6.2, to show that § is adiffeomorphism, it suffices to show that its Jacobian is invertible, i.e., that thegradients V|i- and V\|/• are all linearly independent.

In the following, we first show that the gradients V(j,(- are linearly independent,i.e., that the components of \i can serve as state variables. This result, of course, is notoverly surprising, since it simply says that the system output v and its first r—lderivatives can serve as state variables. We then show that n—r other vector fieldsv|/,- can be found to complete the new state vector, in other words, can be found suchthat the gradients Vu.,- and Vy.- are all linearly independent.

Let us first show that the gradients V(i,- are linearly independent.

Lemma 6.5 If the relative degree of the system (6.67) is r in the region Q, then thegradients V|Xj , V(i.2 , ;.... , V|i r are linearly independent in £l.

Proof: Let us first remark that, in terms of (l, equations (6.71) can be written simply as

Vu,- g = 0 1 < i < r (6.76a)

V|i.r g * o (6.76b)

We present the proof for the case r = 3 (the general case can be shown in a similar fashion).

Assume that there exist smooth functions a;-(x) such that (everywhere in £2)

a3V[i3 = 0 (6.77a)

Multiplying (6.77a) by g yields

[c^Vu, + a 2 V n 2 + a 3 V| i3 ]g = 0

which, from (6.76), implies that 013 = 0 (everywhere in £2).

Replacing a 3 = 0 in (6.77a), in turn, implies that

a ,Vu, + a 2 V u 2 = 0 (6.77b)

Multiplying (6.77b) by the Lie bracket adfg , and using Jacobi's identity and (6.71a), yields

0 = a\Lad%V-\ + <:hLadgV2 =

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (265)

Sect. 6.4 Input-Output Linearization of SISO Systems 251

which, from (6.71b), implies that a 2 = 0 (everywhere in Q).

Replacing a ? = " m (6.77b) implies that

a ,VH[=O (6.77c)

Multiplying (6.77c) by adf2g , and using (6.47) and (6.71a), leads to

0 = o.xLadig\ix = a, [L(2 Lgh -2LtLgLth + LgLf

2 h] = a,LgL f2 h

which, from (6.71b), implies that CC] = 0 (everywhere in £1).

Thus, (6.77a) can only be verified if all the coefficients a(- equal zero everywhere in Q.. This

shows that the gradients V(i; are linearly independent. •

Note that an immediate implication of this result is that, consistently withintuition, the relative degree of an «m-order system is always smaller than n (becausethere cannot be more than n linearly independent vectors in an ^-dimensional space).

Let us now show that there exist n-r more functions \\f: to complete thecoordinate transformation.

Proof: As noticed earlier, given Lemma 6.2, we simply need to show that gradients vectors V\|/-

can be found such that the Vu, and V\|/- are all linearly independent. The development is

illustrated in Figure 6.7.

At xo , equation (6.76a) indicates that the first r— 1 vectors V(i;- (1 = 1, . . . , r— 1 ), which we

have just shown to be linearly independent, are all within the hyperplane orthogonal to g. Since

the dimension of that hyperplane is n— 1 , one can find

n-r = ( n - l ) - ( r - l )

vectors in that hyperplane that are linearly independent of the Vu.,- and linearly independent of

each other. Lei us call these vectors V\j/y (J = 1, . . . , « - / ) . By definition, they verify

Vvj/y- g = 0 1 <j<n~r (6.78)

A subtle point at this stage is that gradient vector fields are not just any vector fields (they must

satisfy curl conditions, cf. also section 3.5.3). Fortunately, Frobenius theorem (Theorem 6.1)

comes to the rescue, since its application to the single vector field g (which is obviously

involutive) guarantees that there does exist n-\ linearly independent gradient functions Vhk

which satisfy Vhk g = 0.

Furthermore, from (6.76b), u r is not in the hyperplane orthogonal to g. Thus the gradients

Vu( (1 = 1, . . . , r) and Vy- (j = 1, . . . , n—r) are all linearly independent. Thus, the Jacobian

of the transformation (6.75) is invertible.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (266)

252 Feedback Linearization Chap. 6

Figure 6.7 : Illustrating the construction

of y

By continuity, the lacobian remains invertible in a neighborhood i2j of x0 . Redefining i i as

the intersection of Qj and £2, the transformation 4/ thus defines a diffeomorphism in Q. Hence, in

Q this transformation is a true state transformation, which puts the nonlinear system into the form

(6.73), with, in (6.73a)

a(n, y) = Lf'h(x) = L(rh[^~

while in (6.73b) the input u indeed does not appear, since from (6.78) the y.y verify

LaMi:(x) = 0 Vxe Q D

From a practical point of view, explicitly finding a vector field y to completethe transformation into a normal form requires the (often nontrivial) step of solvingthe set (6.78) of partial differential equations in the \y-. The following example,adapted from [Isidori, 1989], illustrates this procedure.

Example 6.12: Consider the nonlinear system

2 A"|

> = h(\) = ,v

2 +

e2x2

1/2

(6.79)

Since

y = 2 A'2 = 2 (2 Xj A'2 + sin x^) +

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (267)

Sect. 6.4 Input-Output Linearization of SISO Systems 253

the system has relative degree 2, and

Lfh(x) = 2x2 Lgh{x) = O

LgLth(x)=l

In order to find the normal form, let us take

(i2 = Lf h(x) = 2x2

The third function \|/(x) required to complete the transformation should satisfy

, 3 w -yx 1 3\|/ „L V = —ie2*2+ _ Z = 0

s 3 *! 2dx2

One solution of this equation is

\)/(x) = 1 + JCJ -e 2 j r2

Consider now the associated state transformation z = [ (LLj

V lr- I t s Jacobian matrix is

dx

0 0 l"

0 2 0

1 -2e 2 j r2 0

which is non-singular for any x. The inverse transformation is given by

Thus, this state transformation is valid globally. With the above set of new coordinates, the

system dynamics is put into the normal form

D

THE ZERO-DYNAMICS

By means of input-output linearization, the dynamics of a nonlinear system isdecomposed into an external (input-output) part and an internal ("unobservable") part.Since the external part consists of a linear relation between y and v (or equivalently,

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (268)

254 Feedback Linearization Chap. 6

the controllability canonical form between y and u, as in (6.73a)), it is easy to designthe input v so that the output y behaves as desired. Then, the question is whether theinternal dynamics will also behave well, i.e., whether the internal states will remainbounded. Since the control design must account for the whole dynamics (and thereforecannot tolerate the instability of internal dynamics), the internal behavior has to beaddressed carefully.

The internal dynamics associated with the input-output linearization simplycorresponds to the last (n—r) equations \jf = w(|i, \|/) of the normal form. Generally,this dynamics depends on the output states |i. However, as we now detail, we candefine an intrinsic property of the nonlinear system by considering the system'sinternal dynamics when the control input is such that the output y is maintained at | jzero. Studying this so-called zero-dynamics will allow us to make some conclusionsabout the stability of the internal dynamics.

The constraint that the output y is identically zero implies that all of its timederivatives are zero. Thus, the corresponding internal dynamics of the system, or zero-dynamics, describes motion restricted to the (n-r)-dimensional smooth surface(manifold) Mo defined by \i = 0 . In order for the system to operate in zero-dynamics,i.e., for the state x to stay on the surface Mo , the initial state of the system x(0) mustbe on the surface, and, furthermore, the input u must be such that y stays at zero, i.e.,such that y^(t) = 0 . From (6.69), this means that u must equal

Corresponding to this input, and assuming that indeed the system's initial state is onthe surface, i.e., that ^.(0) = 0, the system dynamics can be simply written in normalform as

|1 = 0 (6.80a)

V = w(0, \f) (6.80b)

By definition, (6.80b) is the zero-dynamics of the nonlinear system (6.67).

The evolution of the system state when operating in zero-dynamics is illustratedin Figure 6.8. Note that, in normal coordinates, the control input u0 can be written as afunction only of the internal states \\i

y)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (269)

Sect. 6.4 Input-Output Linearization ofSISO Systems 255

Example 6.13: Zero-dynamics of a nonlinear system

Consider again the nonlinear system of Example 6.12. Its internal dynamics is represented by theequation

Its zero-dynamics is obtained by letting \il = 0 and (l = 0

The input ug = 0 maintains the output always equal to zero.

Mo

(a) (b)

Figure 6.8 : Evolution of the system state on the zero-dynamics manifold, for n = 3, r- 1.(a) in original state coordinates, (b) in normal coordinates.

Note that computing a system's internal dynamics (or zero-dynamics) does notnecessarily require to first put the system explicitly in normal form, i.e. to solve thep.d.e.'s defining y, which were carefully devised such that u does not appear in y.Indeed, since u (or u0) is known as a function of the state x, it is often easier, given (i,to simply find n—r vector fields p.- to complete the state transformation, and thenreplace u (or uo) by its expression, as we did in section 6.1.3. One can verify that thecorresponding state transformation is one-to-one either directly, or by checking thatthe Jacobian of the transformation is invertible.

Recall from section 6.1.3 that linear systems whose zero-dynamics are stableare called minimum phase. For convenience of reference, we shall extend theterminology to nonlinear systems.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (270)

256 Feedback Linearization Chap. 6

Definition 6.8 The nonlinear system (6.67) is said to be asymptotically minimumphase if its zero-dynamics is asymptotically stable.

The concept of exponentially minimum phase can be defined similarly. If thezero dynamics is asymptotically stable for any y(0) , we shall say the system isglobally asymptotically minimum phase. Otherwise, we shall say the system is locallyminimum phase. For instance, the system (6.79) is globally asymptotically minimumphase.

Let us now consider controller designs based on input-output linearization.Essentially, the idea is to design a controller based on the linear input-output relationand then check whether the internal dynamics is stable or not. In the following, wefirst state a result for local stabilization, and then offer some discussions of globalstabilization and of tracking control.

LOCAL ASYMPTOTIC STABILIZATION

Consider again the nonlinear system (6.67). It is natural to wonder whether choosingthe artificial input v in (6.70) as a simple linear pole-placement controller providesany guarantee about the stability of the overall system. In other words, assume that in(6.70) we let

where the coefficients k[ are chosen such that the polynomial

pr + kr_xpr-x + + klP + k0 (6.81)

has all its roots strictly in the left-half plane. The actual control input u can then bewritten, from (6.68)

u(x) = ^ - [ - L ^ y - kr_xy(>-V - ... - kxy - koy] (6.82)LgL{>

ly

The following result indicates that, provided that the zero-dynamics is asymptoticallystable, this control law indeed stabilizes the whole system locally.

Theorem 6.3 Assume that the system (6.67) has relative degree r, and that its zero-dynamics is locally asymptotically stable. Choose constants kt such that thepolynomial (6.81) has all its roots strictly in the left-half plane. Then, the control law(6.82) yields a locally asymptotically stable closed-loop system.

Proof: The result can be easily understood by writing the closed-loop dynamics in normal

coordinates, where by design

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (271)

Sect. 6.4 Input-Output Linearization ofSISO Systems 257

0 1 0 . . 0"

o o o . . i K = A f l

dd7 V

A O

A, A2

h.o.t.

where h.o.t. denotes higher order terms in the Taylor expansion about xo = 0 , and the matrices

A, A|, and A2 are defined obviously. The above equations can be written

+ h.o.t.

Now since the zero-dynamics is asymptotically stable, its linearization y = A2 y is either

asymptotically stable or marginally stable, according to Lyapunov's linearization theorem

(Theorem 3.1).

If A2 is asymptotically stable (i.e., has all its eigenvalues strictly in the left-half plane), then

the above linearized approximation of the whole dynamics is asymptotically stable, and therefore,

using again Lyapunov's linearization theorem, the nonlinear system is locally asymptotically

stable.

If A2 is only marginally stable, stability of the closed loop system can still be derived by

using the so-called center-manifold theory [Byrnes and Isidori, 1988]. E3

In view of the fact that local stabilization problems may also be solved by linearcontrol methods (through linearization and pole placement), one may wonder aboutthe usefulness of the above theorem. The answer to that question is that the abovestabilization method can treat systems whose linearizations contain uncontrollable butmarginally stable modes, while linear control methods requires the linearized systemto be strictly stabilizable.

For stabilization problems, where state convergence is the objective, one isusually free to choose the output function y = h(x), and thus affect the zero-dynamics.Therefore, it might be possible to choose an output function such that thecorresponding zero-dynamics is asymptotically stable. As a result, the controllerspecified in the above theorem will be able to stabilize the nonlinear system.

Example 6.14: Consider the nonlinear system

=x, x (6.83a)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (272)

258 Feedback Linearization

X2 ~ ^ X2 + u

The system's linearization at x = 0 (where \ = [xl x2] ) is

i,=0

A"o = 3 JCT + M

and thus has an uncontrollable mode corresponding to a pure integrator.

Let us define the output function

Corresponding to this output, the relative degree of the system is 1, because

^ = -2i,-Jc, = -2x,2x,-3jt,-«dt

The associated zero-dynamics (obtained by setting y — 0) is simply

Chap. 6

(6.83b)

(6.84)

and thus is asymptotically stable (cf. Example 3.9). Therefore, according to Theorem 6.3, the

control law

locally stabilizes the nonlinear system.

GLOBAL ASYMPTOTIC STABILIZATION•

As we saw earlier, the stability of the zero-dynamics only guarantees the local stabilityof a control system based on input-output linearization (unless the relative degreeequals the system order, in which case there is no internal dynamics), while thestability of the internal dynamics itself is generally untractable (except in very simplecases, as in section 6.1.3). Thus, it is natural to wonder whether input-outputlinearization ideas can be of any use in (often most practically important) globalstabilization problems. Similarly, one may wonder how to achieve systemstabilization using a given input-output linearization in case the zero-dynamics isunstable.

An approach to global asymptotic stabilization based on partial feedbacklinearization is to simply consider the control problem as a standard Lyapunovcontroller design problem, but simplified by the fact that putting the system in normalform makes part of the dynamics linear. As in most Lyapunov designs based onmathematical rather than physical insights, the approach has a trial-and-error flavor,

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (273)

Sect. 6.4 Input-Output Linearization ofSISO Systems 259

but may actually lead to satisfactory solutions. Let us study it on examples, where thecontrol law based on partial linearization has actually to be modified in order toguarantee the global stability of the whole system.

The basic idea, after putting the system in normal form, is to view n as the"input" to the internal dynamics, and \|/ as its "output". The first step is then to find a"control law" \io = HO(V) which stabilizes the internal dynamics, and an associatedLyapunov function Vo demonstrating that stabilizing property. This is generallyeasier than finding a stabilizing control law for the original system, because theinternal dynamics is of lower order. The second step is then to get back to the originalglobal control problem, define a Lyapunov function candidate V appropriately as amodified version of Vo , and choose the control input v so that V be a Lyapunovfunction for the whole closed-loop dynamics.

Consider, for instance, the problem of stabilizing a nonlinear system whosenormal form is

y = v (6.85a)

z + 'z3+yz = 0 (6.85b)

where v is the control input and \|/ = [z z]T. Considering the internal dynamics(6.85b), we see that if, we had, for instance, y = z2, then the internal dynamics wouldbe asymptotically stable (see Example 3.14). Moreover, this expression of y wouldthen also be compatible with the requirement that y tends to zero. This remarksuggests the following design.

Let Vo be a Lyapunov function indicating the stability of (6.85b) when y is

formally replaced by y0 = z2 . Based on Example 3.14, we can choose

V = ! z 2 + i-z4

° 2 4

Differentiating Vo using the actual dynamics (6.85) leads to

Vo = - z 4 - zz(y-z2)

Consider now the Lyapunov function candidate

V = V +-(y - z 2 ) 2

obtained by adding to Vo a quadratic "error" term in (y - y0). We get

V = -z 4 + (;y-z2)(v-3zz)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (274)

260 Feedback Linearization Chap. 6

This suggest the choice of control law

v = - y + z2 + 3zz

which yields

V = - z 4 - ( y - z 2 ) 2

Application of the invariant set theorem (Theorem 3.5) shows that this choice of vbrings the whole system state to zero.

Interestingly, the same procedure may also be applicable to non-minimumphase systems. Consider, for instance, the problem of stabilizing a nonlinear systemwhose normal form is

y = v

z + z3 - z5 + yz = 0

(6.86a)

(6.86b)

where again v is the control input and y = [ z z]T. Note that the system is non-minimum phase, because its zero dynamics is unstable. Proceeding as before, andnoticing that the internal dynamics would be asymptotically stable if we had y = 2 z4,we let

. 1 66

Differentiating Vo using the actual dynamics leads to

Vo = - z 4 - zz(>>-2z4)

Considering now the Lyapunov function candidate

v= V0+±(y

we get

V = - z 4 + ( j - 2 z 4 ) ( v - 8 z 3 z -z'z)

This suggest the choice of control law

v = ->> + 2z4 + 8z3z + zz

which yields

V = - z 4 - (y-2z4)2

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (275)

Sect. 6.4 Input-Output Linearization ofSISO Systems 261

Application of the invariant set theorem shows that this choice of v brings the wholesystem state to zero.

Another interesting feature of the method is that it can be applied recursively tosystems of increasing complexity, as the following example illustrates.

Example 6.15: Consider the system

y + y'i'-x = 0

z + p -z5 +yz = 0

We first define, of course, a new input v such that

so that the system is in the simpler form

x — v

y + y^z2 = x

z + z 3 -z5 +yz = 0

Consider now only the last two equations and view for a moment x as a control input. Then we

know from our earlier study of the system (6.86) that the "control law" x = x0 where

x . = y-'z2 — y + 2z 4 + 8z 3z + z i

would globally asymptotically stabilize the variables y and z, as can be shown using the

Lyapunov function

" ~ 1Z + 6 Z + 2^y

Consider now the Lyapunov function

2

We have

V=VO + (x-xo)(k-x0) = VO + (x-xo)(v-ko)

= - ! 4 - (y-2z4)2 + (x-xo)(y-2z4 + v~ko)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (276)

262 Feedback Linearization Chap. 6

Define the control law v as

where xg is computed formally from the system dynamics. This yields

V = - z 4 - ( y - 2 z 4 ) 2 - U - * o ) 2

Application of the invariant set theorem then shows that this choice of v brings the whole systemstate to zero. The original control input u can then be computed from v

exists, is bounded, and is uniformly asymptotically stable. Choose constants kt suchthat the polynomial (6.81) has all its roots strictly in the left-half plane. Then, by usingthe control law

u = L — [ - v > i + yd(r) - K-1 % - •••• - W (6-87)

the whole state remains bounded and the tracking error jl converges to zeroexponentially.

The proof of this theorem can be found in [Isidori, 1989]. Note that for trackingcontrol to be exact from time t = 0 on, using any control law, requires that

TRACKING CONTROL

The simple pole-placement controller of Theorem 6.3 can be extended to asymptotictracking control tasks. Consider, for system (6.67), the problem of tracking a givendesired trajectory y^it). Let

and define the tracking error vector by

P(0 = |i(0 - \id(t) If

We then have the following result:I

Theorem 6.4 Assume that the system (6.67) has relative degree r (defined and >constant over the region of interest), that \id is smooth and bounded, and that thesolution \|/rf of the equation

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (277)

Sect. 6.4 Input-Output Linearization of SISO Systems 263

INVERSE DYNAMICS

In the following, we describe the concept of inverse dynamics, which provides aninteresting interpretation of the previous tracking control design based on feedbacklinearization, and also yields insights about the tracking control of non-minimumphase systems.

For systems described by (6.67), let us find out what the initial conditions x(0)and control input u should be in order for the plant output to track a reference outputyr(t) perfectly. To do this, let us assume that the system output y(t) is identical to thereference output yr(t), i.e., y(t)=yr(t) , V r > 0 . This implies that the timederivatives of all orders should be the same as those of the desired output, particularly,

y(k)(t) = y(k\t) k = 0, 1,..., r-1 V t > 0 (6.88)

In terms of normal coordinates, (6.88) can be written

H(0 = \ir(t) = [ yr(t) yr(t) .... yW(t) ]T V t > 0

Thus, the control input u(t) must satisfy

y(-r\t) = a(]in V) + b{\ir, \|/)«(r)

that is,

J r( r ) - a(\i , I)/)

where \|f(?) is the solution of the differential equation

(6-90)

Given a reference trajectory ;y,.(0, we can obtain from (6.89) the required control inputfor output y(t) to be identically equal to yr{i). Note that this output depends on theinternal states \|/(0 and thus, in particular, on the initial y(0).

Normally, by the term "dynamics", we mean the mathematical equations forcomputing the output y(t) of a system corresponding to a given input history u(f).Equations (6.89) and (6.90), on the other hand, allow us to compute the input u{t)corresponding to reference output history yr(t). Therefore, (6.89) and (6.90) are oftencalled the inverse dynamics of the system (6.67). Formally, \|/ is the state of theinverse dynamics, \ir its input, and u its output.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (278)

264 Feedback Linearization Chap. 6

Note from (6.90) that a system's internal dynamics corresponds to inverting thesystem for the reference output yr such that

y/'-) = 3,/-)-£,._! 5K<-!)_ -koy

while its zero-dynamics corresponds to inverting the system for a zero referenceoutput.

TRACKING CONTROL FOR NON-MINIMUM PHASE SYSTEMS

The control law (6.87) cannot be applied to non-minimum phase nonlinear systems,because they cannot be inverted. This is a generalization of the linear result that theinverse of the transfer function of a non-minimum phase linear system is unstable.Therefore, for such systems, we should not look for control laws which achieveperfect or asymptotic convergent tracking errors. Instead, we should find controllerswhich lead to small tracking errors for the desired trajectories of interest.

The control of non-minimum phase systems is a topic of active current research.One interesting approach [Hedrick and Gopalswamy, 1989] is the output-redefinitionmethod, whose principle is to redefine the output function yi = /ij(x) so that theresulting zero-dynamics is stable. Provided that the new output function y\ is definedin such a way that it is essentially the same as the original output function in thefrequency range of interest, exact tracking of the new output function y\ then alsoimplies good tracking of the original output y.

Example 6.16: Tracking control of a non-minimum phase system

To illustrate the idea of redefining the output, let us consider the tracking control of the linear

system

(6.91)A(p)

where b is a strictly positive constant and the zeros of Bo(p) are all in the left-half plane (with p

being the Laplace variable). This system has a right half-plane zero at p = b and, therefore, perfect

tracking and exponentially convergent tracking of an arbitrary desired output yj^t) by y(t) is

impossible. To avoid the problem of unstable zero-dynamics associated with the output y, let us

consider the control of a "nominal output" yi instead, with y^ defined by

'iR t n\

(6.92)

with the desired output for y^ being simply yjj). Based on Theorem 6.4, one can easily find a

control input u that achieves convergent tracking of y^ .

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (279)

Sect. 6.4 Input-Output Linearization of SISO Systems 265

With proper initial conditions, this control law leads to y^(J) =yd{t). What is the tracking

error for the real output? Since one easily sees that the true output y(t) is

thus, the tracking error is proportional to the desired velocity yd ,

and, in particular, is bounded as long as yd is bounded. Interpreting the above in the frequency

domain, we also see that good tracking is achieved when the frequency content of yd is well

below b.

An alternative choice of nominal output is

A(p)(\ + 9b

wich is motivated by the approximation (1 -plb) ~ 1/(1 +p/b) for small \p\/b. With a

feedback linearization design and proper initial conditions, we have y^j) = yjj). This implies that

the true output is

Thus, the tracking error is proportional to the desired acceleration yd,

As compared with the previous nominal output y^, this redefinition allows better tracking if

\yd\ < b \yd\ (or, in the frequency domain, if the frequency content of yd is below b). D

Another practical approximation [Hauser, 1989] may be, when performinginput-output linearization using successive differentiations of the output, to simplyneglect the terms containing the input and keep differentiating the selected output anumber of times equal to the system order, so that there is "approximately" no zero-dynamics. Of course, this approach can only be meaningful if the coefficients of u atthe intermediate steps are "small", i.e., if the systems are "weakly" non-minimumphase. The approach is conceptually similar to neglecting "fast" right-half plane zerosin linear systems (in the frequency domain, 1 — x p ~ 1/(1 + x p) if |xp| « 1 , i.e., ifthe zero (1 /x) is much faster than the frequency range of interest).

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (280)

266 Feedback Linearization Chap. 6

Finally, an approach to dealing with non-minimum phase systems is to modifythe plant itself. In linear systems, while poles can be placed using feedback, zeros areintrinsic properties of the plant and the selected output, and can be changed only bymodifying the plant or the choice of output. Similarly, in nonlinear systems, the zero-dynamics is a property of the plant, the output, and the desired trajectory. It can madestable by changing the output, as discussed earlier. It can also, in principle, bemodified by changing the desired trajectory directly, although this is rarely practical ifthe system is supposed to perform a variety of pre-specified tasks. Finally, it can bemade stable by changing the plant design itself. This may involve relocation oraddition of actuators and sensors, or modifying the physical construction of the plant(e.g., the distribution of control surfaces on an aircraft, or the mass and stiffnessdistributions in a flexible-link robot).

6.5 * Multi-Input Systems

The concepts used in the above sections for SISO systems, such as input-statelinearization, input-output linearization, zero-dynamics, and so on, can be extended toMIMO systems.

In the MIMO case, we consider, in a neighborhood of a point x o , square &systems (i.e., systems with the same number of inputs and outputs) of the form *

Ix = f(x) + G(x)u y = h(x) (6.94) „

*where x is «xl the state vector, u is the mxl control input vector (of components «,-),y is the mxl vector of system outputs (of components >>,•), f and h are smooth vectorfields, and G is a nxtn matrix whose columns are smooth vector fields g (.

FEEDBACK LINEARIZATION OF MIMO SYSTEMS

Input-output linearization of MIMO systems is obtained similarly to the SISO case, bydifferentiating the outputs yi until the inputs appear. Assume that ;,- is the smallestinteger such that at least one of the inputs appears in yfri\ then

with Lg Lf'/"1 ht(x) * 0 for at least one j , in a neighborhood £2, of the point x0 .Performing the above procedure for each output >>(- yields

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (281)

Sect. 6.5 ' Multi-Input Systems 267

Lrfhm(x)

+ E(x)u (6.95)

where the rrtxm matrix E(x) is defined obviously.

Define then Q as the intersection of the Q;-. If, as assumed above, the partial"relative degrees" r,- are all well defined, then Q is itself a finite neighborhood of xo .Furthermore, if E(x) is invertible over the region Q, then, similarly to the SISO case,the input transformation

u =

v _ Jjmh

(6.96)

yields m equations of the simple form

= V: (6.97)

Since the input v,- only affects the output yi, (6.96) is called a decoupling controllaw, and the invertible matrix matrix E(x) is called the decoupling matrix of thesystem. The system (6.94) is then said to have relative degree (rj, . . . . , rm) at xo , andthe scalar r = rj + ... + rm is called the total relative degree of the system at x0 .

An interesting case corresponds to the total relative degree being n. In this case,there is no internal dynamics. With the control law in the form of (6.96), we thusobtain an input-state linearization of the original nonlinear system. With theequivalent inputs v,- designed as in the SISO case, both stabilization and tracking canthen be achieved for the system without any worry about the stability of the internaldynamics. Note that the necessary and sufficient conditions for input-statelinearization of multi-input nonlinear systems to be achievable are similar to and morecomplex than those for single input systems [Su, et al., 1983].

The zero-dynamics of a MIMO system can be defined similarly to the SISOcase, by constraining the outputs to be zero. The notion of a minimum phase systemcan also be similarly defined.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (282)

268 Feedback Linearization

EXTENSIONS OF THE BASIC MIMO LINEARIZATION

Chap. 6

The above input-output linearization can be achieved only when the decoupling matrixE is invertible in the region O. Given the straightforward procedure used to constructE, this condition is rather restrictive (for instance, E may have a column of zeros). Inthe following, we discuss two methods to generate input-output linearization when theinvertibility condition is violated, i.e., when E is singular. Both techniques have aniterative nature, and formal conditions on the system (6.94) can be derived for them toconverge in a finite number of steps (see e.g., [Isidori, 1989], and references therein).The first technique, called dynamic extension, involves choosing some new inputs asthe derivatives of some of the original system inputs, in a way that the correspondingmatrix E is invertible. The control system is designed based on the new set of inputs,and the actual system inputs are then computed by integration. The second technique,a MIMO form of system inversion, involves deriving new outputs so that the resultingE matrix is invertible. In both cases, as in the basic version, the stability of the internaldynamics (or, locally, of the zero-dynamics) has to be verified.

REDEFINING INPUTS: DYNAMIC EXTENSION

For notational simplicity, let us consider a system with 2 inputs and 2 outputs, and letus assume that E(x) has rank one. This means that, without loss of generality, we canredefine the input vector u (through linear transformations) so that E(x) has only onenon-zero column, ej = ej (x), i.e., so that equations (6.95) can be expressed in termsof «! only,

y2(r2) W1 h2(6.98)

Differentiating the above and replacing in the system dynamics then leads to anequation of the form

(6.99)

If the matrix Ej is invertible, then the above equation is in the standard form of (6.95),with «[ and uj regarded as control inputs, and Wj considered as an extra state. Input-output linearization can then be used straightforwardly to design these inputs, i.e.,

A

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (283)

Sect. 6.5 1 Multi-Input Systems 269

(6.100)

with the vector v chosen to place the poles of the resulting linear input-outputdynamics. However, the system input MJ must now be obtained from (6.100) byintegration. Thus the actual control law contains an integrator, yielding a "dynamic"controller.

If, in (6.99), the matrix Ej is still singular, the procedure can be repeated,amounting to adding more integrators.

REDEFINING OUTPUTS: SYSTEM INVERSION

Let us redefine the input as in the dynamic extension case, but now stop at (6.98).Instead of going further to differentiate the left hand side, consider the variable

(where e^=[e^ ex2 ]T)- Using (6.98) shows that the expression of z can becomputed as a function of the state x only (and does not contain the input u), namely

z = en(x) Lf r\ hx(x) - ex x(x) L/i h2{x)

Hence, by differentiating z, we obtain an equation of the form

z = jo(x)+yi(x)ul+j2(x)u2

If the matrix

E2(x) =«ll(x) 0

y,(x) y2(x)

is invertible, we can regard>>] and z as outputs, u\ and u2 as inputs, and use the controllaw

vx~Lxr\hx

to achieve input-output linearization. This leads to

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (284)

270 Feedback Linearization Chap. 6

The new inputs vj and v2 can be easily designed to regulate y and z. If the matrix E2 issingular, we can repeat this procedure to create new outputs.

6.6 Summary

Feedback linearization is based on the idea of transforming nonlinear dynamics into alinear form by using state feedback, with input-state linearization corresponding tocomplete linearization and input-output linearization to partial linearization. Themethod can be used for both stabilization and tracking control problems, single-inputand multiple-input systems, and has been successfully applied to a number of practicalnonlinear control problems, both as an system analysis tool and as a controller designtool.

The method also has a number of important limitations:

1. it cannot be used for all nonlinear systems

2. the full state has to be measured

3. no robustness is guaranteed in the presence of parameter uncertainty orunmodeled dynamics

The applicability of input-state linearization is quantified by a set of somewhatstringent conditions, while input-output feedback linearization cannot be applied whenthe relative degree is not defined and lacks systematic global results. Furthermore,analytically solving the partial differential equations defining input-state linearizingtransformations is generally not systematic.

The second problem is due to the difficulty of finding convergent observers fornonlinear systems and, when an observer can be found, the lack of a generalseparation principle (analogous to that in linear systems) which guarantees that thestraightforward combination of a stable state feedback controller and a stable observerwill guarantee the stability of the closed-loop system.

The third problem is due to the fact that the exact model of the nonlinear systemis not available in performing feedback linearization. The sensitivity to modelingerrors may be particularly severe when the linearizing transformation is poorlyconditioned.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (285)

Sect. 6.8 Exercises 111

Active research is being performed to overcome the above drawbacks. For thefirst problem, research is aimed at extending feedback linearization to non-minimumphase or weakly non-minimum phase systems. For the second problem, many effortsare being made to construct observers for nonlinear systems and to extend theseparation principle to nonlinear systems. For the third problem, robust and adaptivecontrol are being introduced to provide feedback linearizable systems with robustnessto parametric uncertainties. Chapters 7 and 8 provide some further discussions ofrobust and adaptive techniques for feedback linearizable systems.

6.7 Notes and References

Extensive theoretical developments of the material in this chapter can be found in [Isidori, 1989],

from which Examples 6.12 and 6.14 are adapted, and in references therein. The writing of this

chapter was also influenced by the clear presentation in [Spong and Vidyasagar, 1989], and by the

clear and concise summary of input-output linearization in [Hauser, 1989]. The reader may also

consult the recent book [Nijmeijer and Van der Schaft, 1990], which contains many interesting

examples. A survey of aircraft control applications is contained in [Hedrick and Gopalswamy, 1989].

The discussion on global asymptotic stabilization of section 6.4 is inspired by the recent

works of [Kokotovic and Sussmann, 1989; Praly, et al., 1989]. See [Sastry and Kokotovic, 1988;

Sussmann, 1990] for further discussions of the relation between stability of the zero-dynamics and

overall system stability.

While nonlinear observer design is far from being a mature subject, a number of important

theoretical results have been derived [Hermann and Krener, 1977; Vidyasagar, 1980; Krener and

Respondek, 1985]. See, e.g., [Misawa and Hedrick, 1989] for a recent review.

The basic idea of "undoing" the nonlinear dynamics is quite old, and can be traced back at

least to the early robotics literature [Freund, 1973; Takase, 1975]. The basis of input-state

transformations was suggested by [Brockett, 1978] and fully derived in [Jacubczyk and Respondek,

1980; Hunt, et al., 1983]. Many results in input-output linearization are extensions of [Krener, et al.,

1983]. The basic principles of system inversion are due to [Hirschorn, 1981; Singh 1982], and extend

to nonlinear systems the classic linear control algorithm of [Silverman, 1969]. A detailed historical

perspective of the field can be found in [Isidori, 1989].

6.8 Exercises

6.1 Simulate the nonlinear controller design of section 6.1.1 for a spherical tank of radius 0.5

meter (m) and outlet opening radius 0.1 m. Let the initial height be 0.1 m and the desired final

height be 0.6 m.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (286)

272 Feedback Linearization

6.2 Show that the internal dynamics of the system

3

Chap. 6

is unstable.

6.3 For the system

JC> + u

design a controller to track an arbitrary desired trajectory xdl(t) . Assume that the model is very

accurate, that the state [x, x2]T is measured, and that xdl(t), xdl(t), xdl(t) are all known. Write

the full expression of the controller, as a function of the measured state [x, X2]T • Check your

design in simple simulations.

6.4 Consider the system

where the variable xx is to be controlled to zero. Note that the zero-dynamics is exponentially stable.

Show that, assuming that x2(0) = 0, the stability of the internal dynamics depends on the initial

condition in X[ and on how fast u drives Xj to zero. In particular, if these are such that xi > e ' ~ 2/ '11,

then x2 > tan t and therefore tends to infinity in finite time.

6.5 K different interpretation can be given to the zero-dynamics of linear systems. Indeed, note the

frequency-domain equation (6.35) can be written

Ab+bP}uy

Thus, if we assume that the roots of the denominator polynomial are strictly in the left half-plane,

then there are certain control inputs which, after stable transients, do not affect the system output y.

These control inputs verify

J? u = 0h

(6.101)

If the system is minimum-phase, these particular inputs tend to zero as / —> °° . But if the system is

non-minimum phase, then there are certain diverging control inputs which do not affect the system

output (after stable transients). This reflects a "helplessness" of the system input at certain

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (287)

Sect. 6.8 Exercises 273

frequencies.

Show that equation (6.101) is of the same form as that defining the zero-dynamics of the

system.

For the aircraft of Example 6.5, find the control inputs which, after stable transients, do not

affect the output. Illustrate your result in simulation.

Can this interpretation of the zero-dynamics be extended to nonlinear systems?

6.6 Show the bilinearity and skew-commutativity of Lie brackets.

6.7 Check the input-state linearizability of the system

dd7

r vx2xi

X2

Can

23 =

serve as linearizing states? (Adapted from [Nijmeijer and Van der Schaft, 1990].)

6.8 Globally stabilize the nonlinear system

z + z 3 - z 7 + yz2 = 0

where u is the control input. Is the system minimum-phase?

6.9 Put in normal form the nonlinear system

5> + _yz2ln (z4 + 1) = u

z + z5 + z3 + yz2 = 5 u

where u is the control input. Can the system be locally/globally stabilized? Is the system minimum-

phase?

6.10 Consider the system of Figure 6.9, which represents a link driven by an electric motor

through a rigid mechanism, in the vertical plane. The (stable) motor electrical dynamics is assumed

to be slow, so that it cannot be neglected.

Using a procedure similar to that of Example 6.10, perform input-state linearization and

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (288)

274 Feedback Linearization Chap. 6

Motor

Voltage input u

Figure 6.9 : A link driven by a slow motor

design a controller for this system in the following cases (where X is simply a scaling factor between

units):

• The motor dynamics is linear and first-order:

t + XjXx^Xu

• The motor dynamics is linear and second-order:

x + Xlx + ^ 1 , T = X «

• The motor dynamics is nonlinear and second-order:

What variables do you assume to be available for measurement?

In the above calculations, the back-emf of the motor has been neglected. It actually

introduces a damping torque / a0 q . Assuming that the motor dynamics is linear and first-order

(which is often a reasonable model, simply corresponding to an equivalent "RL" circuit), perform

input-state linearization and design a controller for the system. What variables do you assume to be

available for measurement?

6.11 Globally stabilize the nonlinear system

y + z_y4 = w5

z + ( y - l ) z 2 + z5 = 0

where u is the control input. Is the system minimum-phase?

6.12 Consider the nonlinear system

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (289)

Sect. 6.8 Exercises 275

y + z^ey'2 = u

z-(y + y3)('z4+ l) + z5+yz = 0

where u is the control input. Can the system be locally/globally stabilized?

6.13 Globally stabilize the nonlinear system

y+ y^e1 = M3

z - i4 + z5 = y u2z2

where u is the control input. Is the system minimum-phase?

6.14 Globally stabilize the nonlinear system

x + x2y = (x z2 + 4) u

y + y4eyz + x = 0

z + 'zi-z1 +yz2 = 0

where u is the control input. Is the system minimum-phase?

6.15 Discuss the stabilization of the MIMO system

x + xy2 = «j + 2 « 2

y + x 3 = 2 «| + 4 « 2

using dynamic extension and inversion techniques.

Same question for the system

x + xe>' = Mj + 2 «2

5> + y2 = 2 «[ + 4 «2

6.16 Consider again the system of Exercise 6.10 (including motor back-emf) but assume now that

the "link" actually consists of the three blades of an underwater propeller, so that the gravitational

torque - MgLsin q is replaced by the hydrodynamic torque

Xthrust = - / « , 4r |4-|

(where the notation Tlhrml comes from the fact that the thrust generated by the propeller is

proportional to i,hrusl, cf. Exercise 4.4). Assuming that the motor dynamics is linear and first-order,

perform input-state linearization and design a controller for the system. Indicate what variables you

assume to be available for measurement.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (290)

Chapter 7Sliding Control

In this chapter, we consider again the control of nonlinear systems of the general formstudied in chapter 6, but we now allow the models to be imprecise. Modelimprecision may come from actual uncertainty about the plant (e.g., unknown plantparameters), or from the purposeful choice of a simplified representation of thesystem's dynamics (e.g., modeling friction as linear, or neglecting structural modes ina reasonably rigid mechanical system). From a control point of view, modelinginaccuracies can be classified into two major kinds:

• structured (or parametric) uncertainties

• unstructured uncertainties (or unmodeled dynamics)

The first kind corresponds to inaccuracies on the terms actually included in the model,while the second kind corresponds to inaccuracies on (i.e., underestimation of) thesystem order.

As discussed earlier, modeling inaccuracies can have strong adverse effects onnonlinear control systems. Therefore, any practical design must address themexplicitly. Two major and complementary approaches to dealing with modeluncertainty are robust control, which we discuss in this chapter, and adaptive control,which is the subject of chapter 8. The typical structure of a robust controller iscomposed of a nominal part, similar to a feedback linearizing or inverse control law,and of additional terms aimed at dealing with model uncertainty. The structure of anadaptive controller is similar, but in addition the model is actually updated during

276

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (291)

Sect. 7.1 Sliding Surfaces 277

operation, based on the measured performance.

A simple approach to robust control, and the main topic of this chapter, is theso-called sliding control methodology. Intuitively, it is based on the remark that it ismuch easier to control lst-order systems (i.e., systems described by lst-orderdifferential equations), be they nonlinear or uncertain, than it is to control general/ith-order systems (i.e., systems described by «th-order differential equations).Accordingly, a notational simplification is introduced, which, in effect, allowsnth-order problems to be replaced by equivalent lst-order problems. It is then easy toshow that, for the transformed problems, "perfect" performance can in principle beachieved in the presence of arbitrary parameter inaccuracies. Such performance,however, is obtained at the price of extremely high control activity. This is typically atodds with the other source of modeling uncertainty, namely the presence of neglecteddynamics, which the high control activity may excite. This leads us to a modificationof the control laws which, given the admissible control activity, is aimed at achievingan effective trade-off between tracking performance and parametric uncertainty.Furthermore, in some specific applications, particularly those involving the control ofelectric motors, the unmodified control laws can be used directly.

For the class of systems to which it applies, sliding controller design provides asystematic approach to the problem of maintaining stability and consistentperformance in the face of modeling imprecisions. Furthermore, by allowing thetrade-offs between modeling and performance to be quantified in a simple fashion, itcan illuminate the whole design process. Sliding control has been successfully appliedto robot manipulators, underwater vehicles, automotive transmissions and engines,high-performance electric motors, and power systems.

The concepts are presented first for systems with a single control input, whichallows us to develop intuition about the basic aspects of nonlinear controller design.Specifically, section 7.1 introduces the main concepts and notations of sliding control,and illustrates the associated basic controller designs. Section 7.2 describesmodifications of the control laws aimed at eliminating excessive control activity.Section 7.3 discusses the choice of controller design parameters. Section 7.4 thenstudies generalizations to multi-input systems.

7.1 Sliding Surfaces

Consider the single-input dynamic system

x(") =/(x) + b(x) u (7.1)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (292)

278 Sliding Control Chap. 7

where the scalar x is the output of interest (for instance, the position of a mechanicalsystem), the scalar u is the control input (for instance, a motor torque), andx = [JC Jc . . . x(n~^]T is the state vector. In equation (7.1) the function/(x) (ingeneral nonlinear) is not exactly known, but the extent of the imprecision on /(x) is <upper bounded by a known continuous function ofx ; similarly, the control gain b(x) is )not exactly known, but is of known sign and is bounded by known, continuousfunctions of x. For instance, typically, the inertia of a mechanical system is only -sknown to a certain accuracy, and friction models only describe part of the actualfriction forces. The control problem is to get the state x to track a specific time-varying state xrf= [Xrf kd • • • x^"^]^ in the presence of model imprecision onf(x) and b(x).

For the tracking task to be achievable using a finite control u, the initial desiredstate x^(0) must be such that

xd(0) = x(0) (7.2)

In a second-order system, for instance, position or velocity cannot "jump", so that anydesired trajectory feasible from time t = 0 necessarily starts with the same position andvelocity as those of the plant. Otherwise, tracking can only be achieved after atransient.

7.1.1 A Notational Simplification

Let x=x-xdbe the tracking error in the variablex, and let

be the tracking error vector. Furthermore, let us define a time-varying surface Sit) inthe state-space R(") by the scalar equation s(x\t) = 0 , where ?

'in-\

x (7.3) i

and A. is a strictly positive constant, whose choice we shall interpret later. Forinstance, if n = 2 ,

s - x + Xx

i.e., s is simply a weighted sum of the position error and the velocity error; if n = 3 , £

s = x + 2Xx + X^x

I

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (293)

Sect. 7.1 Sliding Surfaces 279

Given initial condition (7.2), the problem of tracking x = xd is equivalent to that ofremaining on the surface S(t)for all t>0 ; indeed s = 0 represents a linear differentialequation whose unique solution is x = 0, given initial conditions (7.2). Thus, theproblem of tracking the n-dimensional vector x^ can be reduced to that of keeping thescalar quantity s at zero.

More precisely, the problem of tracking the ^-dimensional vector xd (i.e., theoriginal «th-order tracking problem in x) can in effect be replaced by a lst-orderstabilization problem in .y. Indeed, since from (7.3) the expression of s containsJ(«-1), we only need to differentiate s once for the input u to appear.

Furthermore, bounds on s can be directly translated into bounds on the trackingerror vector x, and therefore the scalar s represents a true measure of trackingperformance. Specifically, assuming that x(0) = 0 (the effect of non-zero initialconditions in x can be added separately), we have

V t > 0 , | s(t) | < <t> => V t > 0 , 13t«(r) | < (2X)' e (7.4)

i = 0,. . . , n-\

1P + X

1p + X p + X

n -1 blocks

Figure 7.1.a : Computing bounds on x

where e = O / X"~l . Indeed, by definition (7.3), the tracking error x is obtained froms through a sequence of first-order lowpass filters (Figure 7.1.a, where p = (d/dt) is theLaplace operator). Let yl be the output of the first filter. We have

y[(t)= ['e-^ o

From \s\ <O we thus get

\yx(t)\ < O

We can apply the same reasoning to the second filter, and so on, all the way toJn_j = x . We then get

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (294)

280 Sliding Control Chap. 7

| x | < = e

Similarly, JC^ can be thought of as obtained through the sequence of Figure 7.1.b.From the previous result, one has | Zj | < <E>A."~'~' , where zl is the output of thein - i - l) th filter. Furthermore, noting that

p _p + X-X_ j _ Xp+X p + X p + X

P + x

n - i -1 blocks

J V.

1p+X

z Vp+X

-0)

i blocks

Figure 7.1.b : Computing bounds on

one sees that the sequence of Figure 7.1.b implies that

i.e., bounds (7.4). Finally, in the case that x(0)^0, bounds (7.4) are obtainedasymptotically, i.e., within a short time-constant (n - 1)/A,.

Thus, we have in effect replaced an «th-order tracking problem by a lst-orderstabilization problem, and have quantified with (7.4) the correspondingtransformations of performance measures.

The simplified, lst-order problem of keeping the scalar s at zero can now beachieved by choosing the control law u of (7.1) such that outside of Sit)

2 dt(7.5)

where r\ is a strictly positive constant. Essentially, (7.5) states that the squared"distance" to the surface, as measured by s2 , decreases along all system trajectories.Thus, it constrains trajectories to point towards the surface 5(0, as illustrated in Figure7.2. In particular, once on the surface, the system trajectories remain on the surface.In other words, satisfying condition (7.5), or sliding condition, makes the surface aninvariant set. Furthermore, as we shall see, (7.5) also implies that some disturbancesor dynamic uncertainties can be tolerated while still keeping the surface an invariant

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (295)

Sect. 7.1 Sliding Surfaces 281

set. Graphically, this corresponds to the fact that in Figure 7.2 the trajectories off thesurface can "move" while still pointing towards the surface. S(t) verifying (7.5) isreferred to as a sliding surface, and the system's behavior once on the surface is calledsliding regime or sliding mode.

Figure 7.2 : The sliding condition

The other interesting aspect of the invariant set 5(0 is that once on it, the systemtrajectories are defined by the equation of the set itself, namely

n-\

dtx = 0

In other words, the surface S(t) is both a place and a dynamics. This fact is simply thegeometric interpretation of our earlier remark that definition (7.3) allows us, in effect,to replace an «th-order problem by a lst-order one.

Finally, satisfying (7.5) guarantees that if condition (7.2) is not exactly verified,i.e., if x(r=0) is actually off x^t=0), the surface S(t) will nonetheless be reached in afinite time smaller than \s(t=0)\/r\. Indeed, assume for instance that s(t=O) > 0, and let'reach '3e t n e ^me required to hit the surface s = 0. Integrating (7.5) between/ = 0 and t = treach leads to

0-s(t=0) = s(t=tKach) - lv(f=O)<-Ti(/reach-O)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (296)

282 Sliding Control Chap. 7

which implies that

'reach < s(f=O)/T\

One would obtain a similar result starting with s{t=0) < 0 , and thus

rreach < |s(r=O)|/Ti *

Furthermore, definition (7.3) implies that once on the surface, the tracking error tendsexponentially to zero, with a time constant (n - 1)A, (from the sequence of (n - 1)filters of time constants equal to I/A,).

The typical system behavior implied by satisfying sliding condition (7.5) is ''illustrated in Figure 7.3 for n = 2. The sliding surface is a line in the phase plane, ofslope - X and containing the (time-varying) point \d = [xd xrf]

r. Starting from any .initial condition, the state trajectory reaches the time-varying surface in a finite time ijksmaller than |s(?=0)|/r|, and then slides along the surface towards xd exponentially,with a time-constant equal to Ifk.

finite-timereaching phase

sliding modeexponential convergence

Figure 7.3 : Graphical interpretation of equations (7.3) and (7.5) (n = 2)

In summary, the idea behind equations (7.3) and (7.5) is to pick-up a well-behaved function of the tracking error, s, according to (7.3), and then select thefeedback control law u in (7.1) such that s2 remains a Lyapunov-like function of theclosed-loop system, despite the presence of model imprecision and of disturbances.The controller design procedure then consists of two steps. First, as will be illustratedin section 7.1.3, a feedback control law u is selected so as to verify sliding condition(7.5). However, in order to account for the presence of modeling imprecision and of

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (297)

Sect. 7.1 Sliding Surfaces 283

disturbances, the control law has to be discontinuous across S(t). Since theimplementation of the associated control switchings is necessarily imperfect (forinstance, in practice switching is not instantaneous, and the value of s is not knownwith infinite precision), this leads to chattering (Figure 7.4). Now (with a fewimportant exceptions that we shall discuss in section 7.1.4), chattering is undesirablein practice, since it involves high control activity and further may excite high-frequency dynamics neglected in the course of modeling (such as unmodeledstructural modes, neglected time-delays, and so on). Thus, in a second step detailed insection 7.2, the discontinuous control law u is suitably smoothed to achieve an optimaltrade-off between control bandwidth and tracking precision: while the first stepaccounts for parametric uncertainty, the second step achieves robustness to high-frequency unmodeled dynamics.

chattering

Figure 7.4 : Chattering as a result of imperfect control switchings

7.1.2 * Filippov's Construction of the Equivalent Dynamics

The system's motion on the sliding surface can be given an interesting geometricinterpretation, as an "average" of the system's dynamics on both sides of the surface.

The dynamics while in sliding mode can be written as

5 = 0 (7.6)

By solving the above equation formally for the control input, we obtain an expressionfor u called the equivalent control, ue(, , which can be interpreted as the continuous

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (298)

284 Sliding Control Chap. 7

control law that would maintain s = 0 if the dynamics were exactly known. Forinstance, for a system of the form

x=f+ u

we have

and the system dynamics while in sliding mode is, of course,

Geometrically, the equivalent control can be constructed as

ueq = a u+ + (1 - a) u_ (7.7)

i.e., as a convex combination of the values of u on both sides of the surface S(t). Thevalue of a can again be obtained formally from (7.6), which corresponds to requiringthat the system trajectories be tangent to the surface. This intuitive construction issummarized in Figure 7.5, where f+ = [ x f + u+]T , and similarlyf_ = [x f + u_]T and feq = [x f + ueq]. Its formal justification was derived inthe early 1960's by the Russian mathematician A. F. Filippov.

s<0

s>0

Figure 7.5 : Filippov's construction of the equivalent dynamics in sliding mode

Recalling that the sliding motion on the surface corresponds to a limitingbehavior as control switchings occur infinitely fast, the formal solution a of (7.6) and(7.7) can be interpreted as the average "residence time" of the trajectory on the sides>0.

JL

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (299)

Sect. 7.1 Sliding Surfaces 285

7.1.3 Perfect Performance - At a Price

Given the bounds on uncertainties on /(x) and b(x), constructing a control law toverify the sliding condition (7.5) is straightforward, as we now illustrate.

A BASIC EXAMPLE

Consider the second-order system

x=f+u (7.8)

where u is the control input, x is the (scalar) output of interest, and the dynamics /(possibly nonlinear or time-varying) is not exactly known, but estimated a s / . Theestimation error on/is assumed to be bounded by some known function F — F(x,x):

\f~f\<F (7.9)

For instance, given the system

x + a(i) x 2 cos 3x = u (7.10)

where a(t) is unknown but verifies

1 <a(t)<2

one has

/ = -1.5x2cos3x F = 0.5 x2 | cos 3x |

In order to have the system track x(t) s xj^t), we define a sliding surface s = 0according to (7.3), namely:

s = (— + X)x = i + Xx (7.11)

\dt /

We then have:

s = x-x{/+Xx=f+u-x^ + Xx (7.12)

The best approximation u of a continuous control law that would achieve s = 0 is thus

u = -f +xd-l3c (7.13)Note that in terms of the discussion of section 7.1.2, u can be interpreted as our bestestimate of the equivalent control. In order to satisfy sliding condition (7.5) despiteuncertainty on the dynamics / , we add to u a term discontinuous across the surface

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (300)

286 Sliding Control Chap. 7

5 = 0 :

u = u-k&gn(s) (7.14)

where sgn is the sign function:

sgn(s) = + 1 if i > 0sgn(s) = - 1 if s < 0

By choosing k = k(x,x) in (7.14) to be large enough, we can now guarantee that (7.5)is verified. Indeed, we have from (7.12)-(7.14)

s = (f-f)s - k\s\dt

so that, letting

k = F + r\ (7.15)

we get from (7.9)

as desired. Note from (7.15) that the control discontinuity k across the surface s = 0A

increases with the extent of parametric uncertainty. Also, note that/ and F need notdepend only on x or x . They may more generally be functions of any measuredvariables external to system (7.8), and may also depend explicitly on time.

We can see on this basic example one of the main advantages of transformingthe original tracking control problem into a simple first-order stabilization problem ins. Namely, the intuitive feedback control strategy "if the error is negative, push hardenough in the positive direction (and conversely)" actually works for first-ordersystems (recall also Example 3.9). It does not for higher-order systems.

INTEGRAL CONTROL

A similar result would be obtained by using integral control, i.e., formally letting([' x(r) dr) be the variable of interest. The system (7.8) is now third-order relative to

this variable, and (7.3) gives:

We then obtain, instead of (7.13),

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (301)

Sect. 7.1 Sliding Surfaces 287

u = —f + xj — 2 A, x — 7? x

with (7.14) and (7.15) formally unchanged. Note that f'xdr can be replaced by

f' x dr, i.e., the integral can be defined to within a constant. The constant can bechosen to obtain s(t=0) = 0 regardless of x^(0), by letting

: = 2 + 2Xx + X2j'x dr - 3?(0)

GAIN MARGINS

Assume now that (7.8) is replaced by

x=f+bu (7.16)

where the (possibly time-varying or state-dependent) control gain b is unknown but ofknown bounds (themselves possibly time-varying or state-dependent)

0 < bmin <b<bmax (7.17)

Since the control input enters multiplicatively in the dynamics, it is natural to chooseA

our estimate b of gain b as the geometric mean of the above bounds:

b = (bmjnbmax)l/2

Bounds (7.17) can then be written in the formA

P"'< £ <p (7.18)b

where

Since the control law will be designed to be robust to the bounded multiplicativeuncertainty (7.18), we shall call (3 the gain margin of our design, by analogy to theterminology used in linear control. Note that (3 may be time-varying or state-dependent, and that we also have

Also note that the uncertainty on b may come directly in the form (7.18), e.g., if thecontrol action u itself is generated by a linear dynamic system.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (302)

288 Sliding Control Chap. 7

With 5 and u defined as before, one can then easily show that the control law

u = b~l[u-ksgn(s)] (7.19)

with

k > P(F + iD + ( p - l ) | « | (7.20)

satisfies the sliding condition. Indeed, using (7.19) in the expression of s leads to

s = (f- bb~l / ) + (1 - btrx){-xd + Ai) - bb~lk sgn(s)

so that k must verify

k > | bb-lf-f + (bb-{ - l)(-xd + Xk) | + r ^ " 1

A A A

Since/=/ + ( / - / ), where \f-f I < F, this in turn leads to

b

k > bb~x F + r\bb~: + \bb~l -l\-\f-xd +

and thus to (7.20). Note that the control discontinuity has been increased in order to

account for the uncertainty on the control gain b.

Example 7.1: A simplified model of the motion of an underwater vehicle can be written

mx + cx{x\ = u (7.21)

where x defines position, w is the control input (the force provided by a propeller), m is the mass

of the vehicle (including the so-called added-mass, associated with motion in a fluid), and c is a

drag coefficient. In practice, m and c are not known accurately, because they only describe

loosely the complex hydrodynamic effects that govern the vehicle's motion.

Defining s as s = x + Xx , computing i explicitly, and proceeding as before, a control law

satisfying the sliding condition can be derived as

u = m( Xj -Xx) + cx\x\-k sgn i (7.22)

with

k = (F+$r\) + m{$-l)\xd-lJc\ (7.23)

Note that expression (7.23) is "tighter" than the general form (7.20), reflecting the simpler

structure of parametric uncertainty: intuitively, u can compensate for c x | x | directly, regardless

of the uncertainty on m. In general, for a given problem, it is a good idea to quickly rederive a

control law satisfying the sliding condition, rather than apply some pre-packed formula. D

It is useful to pause at this point, and wonder whether a different control action,

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (303)

Sect. 7.1 Sliding Surfaces 289

obtained by some other method, could achieve the same task. The answer to thisquestion is that, given a feasible desired trajectory, there is a unique smooth controltime-function that tracks it exactly, namely

[xd-f(xd)] (7.24)

Thus, whatever the method, the resulting control time-function will be the same, andtherefore using this particular approach simply provides a straightforward way ofarriving at that time-function. Because we require perfect tracking to be achieved evenin the presence of parametric uncertainty, this time-function is constructed through aprocess of averaging infinitely fast discontinuous switchings, into what we called insection 7.1.2 the equivalent control, which is precisely (7.24).

Control laws which satisfy sliding condition (7.5), and thus lead to "perfect"tracking in the face of model uncertainty, are discontinuous across the surface S(t),thus leading in practice to control chattering. In general, chattering is highlyundesirable, since it involves extremely high control activity, and furthermore mayexcite high-frequency dynamics neglected in the course of modeling. In section 7.2,we shall show how to modify the switching control laws derived above so as toeliminate chattering.

In specific (if exceptional) applications, however, control chattering isacceptable, and the pure switching control laws derived above can yield extremelyhigh performance. We now discuss such direct applications of the previousdevelopment.

7.1.4 Direct Implementations of Switching Control Laws

The main direct applications of the above switching controllers include the control ofelectric motors, and the use of artificial dither to reduce stiction effects.

SWITCHING CONTROL IN PLACE OF PULSE-WIDTH MODULATION

In pulse-width modulated electric motors, the control input u is an electrical voltagerather than a mechanical force or acceleration. Control chattering may then beacceptable provided it is beyond the frequency range of the relevant unmodeleddynamics. Provided that the necessary computations (including both control law andstate estimation) can be handled on-line at a high enough rate, or implemented usinganalog circuitry, pure sliding mode control using switched control laws can be a viableand extremely high-performance option.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (304)

290 Sliding Control Chap. 7

SWITCHING CONTROL WITH LINEAR OBSERVER

The difficulty in obtaining meaningful state measurements at very high sampling ratescan be turned around by using state observers. For linear systems, the design of suchobservers is well known and systematic. The principle of the approach to designing aswitching controller using an observer is then very simple. Instead of tracking thesurface 5 = 0, the system is made to track the surface se = 0, where se is obtained byreplacing the state x by its estimate xe in the expression of s. This can be achieved bycomputing a dynamic compensation term ue based on the available state estimates,and using switching terms of the form - k(xe) sgn(se), where k(xe) is chosen largeenough to compensate both for parametric inaccuracies and for observer inaccuracies.This yields se -> 0 (as t —> °°). Then, if the observer has been properly designed sothat it converges despite modeling uncertainties (which, again, is easy to achieve inthe linear case), we also have s —> se . Therefore, s —» 0, and the actual stateconverges towards the desired state. Furthermore, sliding mode and its robustproperties are maintained on the surface se = 0, which tends towards the desiredsliding surface as the observer converges.

SWITCHING CONTROL IN PLACE OF DITHER

When uncertainty consists of effects small in magnitude but difficult to model, such asstiction or actuator ripple, switching in s may be advantageously used in place of amore standard "dither" signal. Ideally, the frequency of the switching should bechosen well beyond that of significant structural vibration modes (in mechanicalsystems), while remaining below the actuators' bandwidth. This assumes again thatmeaningful state estimates can be provided at the selected switching frequency. Suchan approach can particularly improve the quality of low-speed behavior, whichotherwise is extremely sensitive to friction.

The examples above represent the few specific applications where chatteringcan be tolerated and actually exploited. In the general case, however, the question ishow to derive control laws that maintain the system close to the surface 5 = 0 whileavoiding chattering altogether. This is the subject of the next section.

7.2 Continuous Approximations of Switching Control Laws

In general, chattering must be eliminated for the controller to perform properly. Thiscan be achieved by smoothing out the control discontinuity in a thin boundary layerneighboring the switching surface

B{t) = (x, Kx;0l<<&) $ > 0 (7.25)

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (305)

Sect. 7.2 Continuous Approximations of Switching Control Laws 291

s = 0

Figure 7.6.a : The boundary layer

where O is the boundary layer thickness, and e = <$>ftin~l is the boundary layer width,as Figure 7.6.a illustrates for the case n = 2. In other words, outside of B(t), we choosecontrol law u as before {i.e., satisfying sliding condition (7.5)), which guarantees thatthe boundary layer is attractive, hence invariant: all trajectories starting inside B(t=0)remain inside B(t) for all r > 0 ; and we then interpolate u inside B(t) - for instance,replacing in the expression of u the term sgn(j) by s/<3>, inside B(t), as illustrated inFigure 7.6.b.

Given the results of section 7.1.1, this leads to tracking to within a guaranteedprecision e (rather than "perfect" tracking), and more generally guarantees that for alltrajectories starting inside B(t=0)

V t > 0 , | < (2X)' e i = 0, ... , n - l

Example 7.2: Consider again the system (7.10), and assume that the desired trajectory isxd= sin(;ir/2).

Figure 7.7 shows the tracking error and control law using the switched control law (withA 20 01)

u = u~k sgn(.s)

= 1.5 x2 cos 3x + xd - 20 1 - ( 0.5 x2 | cos 3x | + 0.1) sgn[ j + 20 x]

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (306)

292 Sliding Control Chap. 7

Figure 7.6.b : Control interpolation in the boundary layer

at a sampling rate of 1 kHz. The actual value of a(t) used in the simulations is a(f) = |sinf| + 1

(which verifies the assumed bound on a(t)). We see that tracking performance is excellent, but is

obtained at the price of high control chattering.

Assume now that we interpolate the above control input in a thin boundary layer of thickness

0.1

u = u - k sat(s/<t>)

= 1.5 x2 cos 3x + xd- 20 i - (0.5 x2 | cos 3x | + 0.1) sat[(j? + 20 x)/0.1 ]

As shown in Figure 7.8, the tracking perfomiance, while not as "perfect" as above, is still very

good, and is now achieved using a smooth control law. Note that the bounds on tracking error are

consistent with (7.25). D

6.0

4.0

2.0

0.0

-2.0

-4.00.0 1.0 2.0 3.0 4.0

time(sec)3.0 4.0time(sec)

Figure 7.7 : Switched control input and resulting tracking performance

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (307)

Sect. 7.2 Continuous Approximations of Switching Control Laws 293

6.0

4.0

2.0

0.0

-2.0

-4.0

4e-03

0.0 1.0 2.0 3.0 4.0time(sec)

-4e-030.0 1.0 2.0 3.0 4.0

time(sec)

Figure 7.8 : Smooth control input and resulting tracking performance

The intuitive understanding of the effect of control interpolation in a boundarylayer can be carried on further, and guide the selection of the design parameters A, and<I>. As we now show, the smoothing of control discontinuity inside B(t) essentiallyassigns a lowpass filter structure to the local dynamics of the variable s, thuseliminating chattering. Recognizing this filter-like structure then allows us, inessence, to tune up the control law so as to achieve a trade-off between trackingprecision and robustness to unmodeled dynamics. Boundary layer thickness <B can bemade time-varying, and can be monitored so as to well exploit the control "bandwidth"available. The development is first detailed for the case p = 1 (no gain margin), andthen generalized.

A

Consider again the system (7.1) with b = b=l. In order to maintainattractiveness of the boundary layer now that <P is allowed to vary with time, we mustactually modify condition (7.5). Indeed, we now need to guarantee that the distance tothe boundary layer always decreases

=> _ [ J -_ddf'

< -r\

S<-<P => —\_S- {-<!>)[ > T\dt

Thus, instead of simply requiring that (7.5) be satisfied outside the boundary layer, wenow require that (combining the above equations)

=> I £ J 2 < ( O -2 dt

(7.26)

The additional term O |$| in (7.26) reflects the fact that the boundary layer attractioncondition is more stringent during boundary layer contraction (O < 0) and lessstringent during boundary layer expansion (<t> > 0). In order to satisfy (7.26), the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (308)

294 Sliding Control Chap. 7

quantity -<D is added to control discontinuity gain fc(x) , i.e., in our smoothedimplementation the term k(\) sgn(s) obtained from switched control law u is actually

— f 1replaced by k (x) sat(s/<t>), where

k (x) = fc(x) - <t> (7.27)

and sat is the saturation function, which can be formally defined as

sat(>>) = y if\y\<l

sat(y) = sgn(>>) otherwise

Accordingly, control law u becomes:

u - u - k (x) sat(s/<l>)

Let us now consider the system trajectories inside the boundary layer, wherethey lie by construction: they can be expressed directly in terms of the variable s as

S = -F(x) ^ - A/(x) (7.28)

AA

where A / = / - / . Now since k and A/are continuous in x, we can exploit (7.4) torewrite (7.28) in the form

s = - ~k(xd) ^ + ( - A/(xd) + O(e)) (7.29)

We see from (7.29) that the variable s (which is a measure of the algebraic distance tothe surface S(t)) can be viewed as the output of a first-order filter, whose dynamicsonly depends on the desired state xJJ), and whose inputs are, to the first order,"perturbations," i.e., uncertainty A/(x^). Thus, chattering can indeed be eliminated, aslong as high-frequency unmodeled dynamics are not excited. Conceptually, thestructure of the closed-loop error dynamics can be summarized by Figure 7.9:perturbations are filtered according to (7.29) to give s, which in turn provides trackingerror x by further lowpass filtering, according to definition (7.3). Control action u is afunction of x and xd. Now, since X is the break-frequency of filter (7.3), it must bechosen to be "small" with respect to high-frequency unmodeled dynamics (such asunmodeled structural modes or neglected time delays). Furthermore, we can now tunethe boundary layer thickness <& so that (7.29) also represents a first-order filter of

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (309)

Sect. 7.2 Continuous Approximations of Switching Control Laws 295

bandwidth L It suffices to let

which can be written from (7.27) as

<T> + A.O = k(xd)

(7.30)

(7.31)

1 order filter(7-29)

1

(P+X)"-1

CHOICE OF <P DEFINITION OF s

Figure 7.9 : Structure of the closed-loop error dynamics

Equation (7.31) defines the desired time-history of boundary layer thickness <J>, and, inthe light of Figure 7.9, shall be referred to as the balance condition. Intuitively, itamounts to tuning up the closed-loop system so that it mimics an «tn order criticallydamped system. Furthermore, definition (7.27) can then be rewritten as

k (x) = k(x) - k(xd) + X® (7.32)

The s-trajectoiy, i.e., the variation of s with time, is a compact descriptor of theclosed-loop behavior: control activity directly depends on s, while by definition (7.3)tracking error x is merely a filtered version of s. Furthermore, the s-trajectoryrepresents a time-varying measure of the validity of the assumptions on modeluncertainty. Similarly, the boundary layer thickness <J> describes the evolution ofdynamic model uncertainty with time. It is thus particularly informative to plots(t), ®(t), and - <t>(r) on a single diagram, as illustrated in Figure 7.10.

Example 7.3: Consider again the system described by (7.10). The complete control law is now

u = xd-Xx + 1.5 x2 cos 3x - (0.5 x21 cos 3x | + r| - 4>) sat[(? + Xx)/<t>]

with 4> = - X$> + (0.5 xd2 | cos 3xd | +1|)

and, assuming e.g., that xjifi) = 0 initially, 0(0) = r\/X . As in Example 7.2, we let T] = 0.1 and

X — 20 . Typically, the arbitrary constant T| (which, formally, reflects the time to reach the

boundary layer starting from the outside) is chosen to be small as compared to the average value

of k(\d), so as to fully exploit our knowledge of the structure of parametric uncertainty. The

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (310)

296 Sliding Control Chap. 7

Figure 7.10 : The ^-trajectories can convey much information on a single plot

value of X is selected based on the frequency range of unmodeled dynamics, as we shall discuss

later.

The tracking error, control input, and ^-trajectories are plotted in Figure 7.11 for the same

desired trajectory xd = sin(7U/2) as in Example 7.2. We see that while the maximum value of the

time-varying boundary layer thickness <J> is the same as that originally chosen (purposefully) as

the constant value of O in Example 7.2, the tracking error is consistently better (up to 4 times

better) than that in Example 7.2, because varying the thickness of the boundary layer allows us to

make better use of the available bandwidth. C\

1.0 2.0 3.0 4.0time(sec)

1.0 2.0 3.0 4.0time(sec)

Figure 7.11a : Control input and resulting tracking performance

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (311)

Sect. 7.2 Continuous Approximations of Switching Control Laws 297

Figure 7.11b : ^-trajectories with time-varying boundary layers

In the case that p •£ 1, one can easily show that (7.31) and (7.32) become (with

=> (7.33)

=> Piwith initial condition 0(0) defined as:

Indeed, in order to satisfy (7.26) in the presence of uncertainty (3 on the control gain we let

6 => k~(x) = k(x) - O/p

=> T(x) = *(x) - P 4>

Furthermore, the balance condition can be written, instead of (7.30), as

(7.34)

(7.35)

(7.36)

(7.37)

(7.38)

that is,

k (xd) =

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (312)

298 Sliding Control Chap. 7 fj

Applying this relation to (7.36), (7.37) leads to the desired behavior of <I>:

=> ^Vd

which we can combine with (7.36)-(7.37) and rewrite as (7.33)-(7.34). Finally, remark that if

(3 = (5 , one has

k (x) = (k (x) - k (xrf)) + k (xd) = k(x) - k(xd)

Note that the balance conditions (7.33) and (7.34) imply that <J> and thus x arebounded for bounded xd.

The balance conditions have a simple and intuitive physical interpretation:neglecting time constants of order I/A., they imply that

that is

(bandwidth)" x (tracking precision)= (parametric uncertainty measured along the desired trajectory)

Such trade-off is quite similar to the situation encountered in linear time-invariantsystems, but here applies to the more general nonlinear system structure (7.1), allalong the desired trajectory. In particular, it shows that, in essence, the balanceconditions specify the best tracking performance attainable, given the desired controlbandwidth and the extent of parameter uncertainty. Similarly, expression (7.20)allows one to easily quantify the trade-off between accuracy and speed along a givenpath, for instance.

Example 7.4: Let us consider again the underwater vehicle of Example 7.1, and smoothen the

control input using time-varying boundary layer, as described above. The a priori bounds on m

and c are

1 < m < 5 0.5<c<1.5

and, accordingly,

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (313)

Sect. 7.2 Continuous Approximations of Switching Control Laws 299

The actual values used in the simulation are

1.5sin(liU) c= 1.2 + .2sin(lilr)

which are used as a metaphor of the complexity of the actual hydrodynamic effects. We let

i^=0.1 and X = 10.

The desired trajectory consists of a constant-acceleration phase at 2 m/s2 for two seconds, a

constant-velocity phase (at 4 m/s) for two seconds, and a constant-acceleration phase at - 2 m/s2

for two seconds. The corresponding tracking error, control input, and s-trajectories are plotted in

Figure 7.12. •

§

0 -

-100.0 1.0 2.0 3.0 4.0 5.0 6.0

time(sec)0.0 1.0 2.0 3.0 4.0 5.0 6.0

time(sec)

Figure 7.12a : Control input and resulting tracking performance

0.0 1.0 2.0 4.0 5.0 6.0timefsecl

Figure 7.12b : ^-trajectories with time-varying boundary layers

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (314)

300 Sliding Control

REMARKS

Chap. 7

(i) The desired trajectory xd must itself be chosen smooth enough not to excitethe high-frequency unmodeled dynamics.

(ii) An argument similar to that of the above discussion shows that the choice ofdynamics (7.3) used to define sliding surfaces is the "best-conditioned" among lineardynamics, in the sense that it guarantees the best tracking performance given thedesired control bandwidth and the extent of parameter uncertainty.

(iii) If the model or its bounds are so imprecise that F can only be chosen as a

large constant, then <j> from (7.31) is constant and large, so that the term k sat(s/(j>)simply equals X s/P in the boundary layer, and therefore acts as a simple P.D.: there isno free lunch.

(iv) A well-designed controller should be capable of gracefully handlingexceptional disturbances, i.e., disturbances of intensity higher than the predictedbounds which are used in the derivation of the control law. For instance, somebodymay walk into the laboratory and push violently on the system "to see how stiff it is";an industrial robot may get jammed by the failure of some other machine; an actuatormay saturate as the result of the specification of an unfeasible desired trajectory. Ifintegral control is used in such cases, the integral term in the control action maybecome unreasonably large, so that once the disturbance stops, the system goesthrough large amplitude oscillations in order to return to the desired trajectory. Thisphenomenon, known as integrator windup, is a potential cause of instability becauseof saturation effects and physical limits on the motion. It can be simply avoided bystopping integration (i.e. maintaining the integral term constant) as long as the systemis outside the boundary layer. Indeed, under normal circumstances the system doesremain in the boundary layer; on the other hand, when the conditions return to normalafter an exceptional disturbance, integration can resume as soon as the system is backin the boundary layer, since the integral term is defined to within an arbitrary constant.

(v) In the case that X is time-varying (as further discussed in the next section),the term

u =-Xx

should be added to the corresponding u, while augmenting gain k{\) accordingly bythe quantity | u |((3 - 1).

The degree of simplification in the system model may be varied according tothe on-line computing power available: in essence, the balance conditions quantify the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (315)

Sect. 7.3 The Modeling/Performance Trade-Offs 301

trade-off between model precision and tracking accuracy, as further detailed next.Furthermore, the ^--trajectories provide a measure of the validity of the assumptionson model uncertainty and of the adequacy of bound simplifications.

7.3 The Modeling/Performance Trade-Offs

The balance conditions (7.33)-(7.34) have practical implications in terms ofdesign/modeling/performance trade-offs. Neglecting time-constants of order I/A.,conditions (7.33) and (7.34) imply that

(7.39)

as noticed in section 7.2. If we now consider the structure of control law (7.19), wesee that the effects of parameter uncertainty on / have been "dumped" in gain k.Conversely, better knowledge of/reduces k by a comparable quantity. Thus (7.39) isparticularly useful in an incremental mode, i.e., to evaluate the effects of modelsimplification (or conversely of increased model sophistication) on trackingperformance:

(7.40)

In particular, marginal gains in performance are critically dependent on controlbandwidth X: if large X's are available, poor dynamic models may lead to respectabletracking performance, and conversely large modeling efforts produce only minorabsolute improvements in tracking accuracy.

It is of course not overly surprising that system performance be very sensitive tocontrol bandwidth A. : (7.1) only represents part of the system dynamics - e.g., itsrigid-body component - while X accounts for the unmodeled part. In the right-handside of (7.40), the effects of parametric uncertainty in (7.1) are reflected in thenumerator, while the presence of dynamics neglected in the model is reflected in thedenominator, since it both translates into the order n of the model, and imposes upperbounds on the choice of X.

Thus, given model (7.1), a key question is to determine how large X can bechosen. Although the tuning of this single scalar may in practice be doneexperimentally, considerable insight on the overall design can be obtained byexplicitly analyzing the various factors limiting X. In mechanical systems, forinstance, given clean measurements, X is typically limited by three factors:

(i) structural resonant modes: X must be smaller than the frequency Vg of thelowest unmodeled structural resonant mode; a reasonable interpretation of this

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (316)

302 Sliding Control

constraint is, classically

- ^-R °* -j VR

Chap. 7

(7.41)

although in practice this bound may be modulated by engineering judgment, takingnotably into account the natural damping of the structural modes. Furthermore, itmay be worthwhile in certain cases to account for the fact that XR may actuallyvary with the task (e.g., given configuration or loads).

(ii) neglected time delays: along the same lines, we have a condition of theform

J (7.42)

when T^ is the largest unmodeled time-delay (for instance in the actuators).

(iii) sampling rate: with a full-period processing delay, one gets a condition ofthe form

^sampling (7.43)

where vsam lin is the sampling rate.

The desired control bandwidth X is the minimum of the three bounds (7.41)-(7.43).Bound (7.41) essentially depends on the system's mechanical properties, while (7.42)reflects limitations on the actuators, and (7.43) accounts for the available computingpower. Ideally, the most effective design corresponds to matching these limitations,i.e., having

«XA^XS = X (7.44)

Now (7.41) and (7.42) are "hard" limitations, in the sense that they representproperties of the hardware itself, while (7.43) is "soft" as far as it reflects theperformance of the computer environment and the complexity of the controlalgorithm. Assume for instance that bound (7.43) is the most stringent, which meansthat the system's mechanical potentials are not fully exploited. This may typicallyoccur in modern high-performance robots (such as direct-drive arms) which featurehigh mechanical stiffness and high resonant frequencies. It may be worthwhile, beforeembarking in the development of dedicated computer architectures, to first considersimplifying the dynamic model used in the control algorithm. This in turn allows oneto replace X = XS\OVJ by a larger X = A.fast which varies inversely proportionally to the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (317)

Sect. 7.4 * Multi-Input Systems 303

required computation time. From (7.40) and assuming that neither of bounds (7.41) or(7.42) is hit in the process, this operation is beneficial as long as

_ , ( 7 . 4 5 )

Conversely, equality in (7.45) defines the threshold at which model simplificationstarts degrading performance despite gains in sampling rate. This threshold is rarelyreached in practice: even assuming that marginal gains in model precision dependlinearly on the computation time involved, X~2 still varies as the square of therequired sampling period. Thus it is often advisable to reduce model complexity untilcomputation can be achieved at a rate fully compatible with the mechanicalcapabilities of the arm, in other words until \$ is no longer the "active" limit on X.The performance increase resulting from this simple operation may in turn beadequate and avoid major development efforts on the computer hardware.

The trade-off between modeling inaccuracy and performance can be furtherimproved only by updating the model on-line. This can indeed be achieved whensome components of the model depend linearly on unknown but constant parameters,allowing the corresponding uncertainties to be mapped in terms of a single unknownconstant vector. This is the topic of chapter 8, adaptive control.

7.4 * Multi-Input Systems

This section discusses extensions of the previous development to multi-input systems.The point of view taken here is essentially mathematical. In chapter 9, we shall discusshow the exploitation of known physical properties of the systems, such asconservation of energy, may often make such multi-input extensions simpler and morepowerful.

Consider a nonlinear multi-input system of the form

m

x i { r t i ) = - / / W + X b i j W uj i = l , . . . , m j = 1 , . . ., mi= i

where the vector u of components u-. is the control input vector, and the state x iscomposed of the xfs and their first («;- - 1) derivatives. As mentioned in chapter 6,such systems are called square systems, since they have as many control inputs «.• asoutputs to be controlled Xj. We are interested again in the problem of having the statex track a desired time-varying state xd , in the presence of parametric uncertainties.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (318)

304 Sliding Control Chap. 7

We make two assumptions. First, we assume that the matching conditionsdiscussed in chapter 6 are verified, i.e., that parametric uncertainties are within therange space of the input matrix B (of components btj). Since B is a square mxmmatrix, this simply means that B is invertible over the whole state-space, acontrollability-like assumption. Second, we assume that the estimated input matrixA A

B is invertible, continuously dependent on parametric uncertainty, and such that B = Bin the absence of parametric uncertainty.

As in the single-input case, we shall write uncertainties on f in additive form,and uncertainties on the input matrix B in multiplicative form

\}'i~fi\<Fi i = l , . . . , m (7.46)

B = (I + A) B i=l, ...,m 7 = 1, ...,m (7.47)

where I is the n x n identity matrix. Note that the structure of expression (7.46) isslightly different from that of (7.47), since the notion of a gain margin is mostly ascalar concept, while (7.47) shall prove more convenient for the purpose of matrixmanipulation.

Let us define a vector s of components st by

n:-\

which, for notational compactness, we shall write

This defines a vector x/"~ ^ of components xrfni ~ , which can be computed from x

and xd . As in the single-input case, the controller design can be translated in terms offinding a control law for the vector u that verifies individual sliding conditions of theform

1 d ,2 < _i'1 ~ (7.48)

in the presence of parametric uncertainty. Letting k sgn(s) be the vector ofcomponents kv sgn(s(), and choosing the control law to be of the form

u = B" J ( x / " - ' ) - f - k sgn(s))

similarly to the single-input case, we can write

(7.49)

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (319)

Sect. 7.4 * Multi-Input Systems 305

h = ft-fi + SAW"'--0-/,-)- ] £ A,y £y sgn(.yy) - (I + A/7) ^-sgn(^)

j * •

Thus, the sliding conditions are verified if

> = 1 J• * i

/ = 1 , . . . ,«

and, in particular, if the vector k is chosen such that

(1 - Du) ki + J Dij kj = F{ + j^Dy | xrf"i - I) - /,-1 + n,- (7.50)y * l" y = 1

/ = 1, . . . ,n

Expression (7.50) represents a set of m equations in the m switching gains kj.Do these equations have a solution k (then necessarily unique), and are thecomponents kj all positive (or zero)? The answer to both questions is yes, thanks to aninteresting result of matrix algebra, known as the Frobenius-Perron theorem.

Theorem (Frobenius-Perron) Consider a square matrix A with non-negativeelements. Then, the largest real eigenvalue p± of A of is non-negative. Furthermore,consider the equation

(I - p - 1 A) y = z

where all components of the vector z are non-negative. If p > pj , then the aboveequation admits a unique solution y, whose components yt are all non-negative.

Applying the Frobenius-Perron theorem to the matrix of components D,-., and noticingthat our second assumption on the system implies that Pi < 1, shows that equation(7.50) uniquely defines a set of non-negative kj. Thus, the control law (7.49), with kdefined by (7.50), satisfies the sliding condition in the presence of parametricuncertainties bounded by (7.46).

As in the single-input case, the switching control laws derived above can besmoothly interpolated in boundary layers, so as to eliminate chattering, thus leading toa trade-off between parametric uncertainty and performance. The reader is referred tosection 7.6 for details.

Note that the point of view taken in this section is essentially mathematical.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (320)

306 Sliding Control Chap. 7

Chapter 9 shall dvscuss how the exploitation of physical properties of the systems,such as conservation of energy, often makes multi-input designs simpler and morepowerful. This will become particularly important in adaptive versions of the designs.

7.5 Summary

The aim of a sliding controller is to

(i) Design a control law to effectively account for

• parameter uncertainty, e.g., imprecision on the mass properties orloads, inaccuracies on the torque constants of the actuators, friction,and so on.

• the presence of unmodeled dynamics, such as structural resonantmodes, neglected time-delays (in the actuators, for instance), or finitesampling rate.

(ii) Quantify the resulting modeling!performance trade-offs, and inparticular, the effect on tracking performance of discarding any particularterm in the dynamic model.

The methodology is based on a notational simplification, which amounts toreplacing an nth order tracking problem by a first order stabilization problem.Although "perfect" performance can in principle be achieved in the presence ofarbitrary parameter inaccuracies, uncertainties in the model structure (i.e., unmodeleddynamics) lead to a trade-off between tracking performance and parametricuncertainty, given the available "bandwidth." In practice, this corresponds to replacinga switching, chattering control law by its smooth approximation. In specificapplications where control chattering is acceptable, the pure switching control lawscan yield extremely high performance.

Sliding controller design provides a systematic approach to the problem ofmaintaining stability in the face of modeling imprecisions. Furthermore, it quantifiesthe modeling/performance trade-offs, and in that sense can illuminate the wholedesign and testing process. Finally, it offers the potential of simplifying higher-levelprogramming, by accepting reduced information about both task and system.

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (321)

Sect. 7.7 Exercises 307

7.6 Notes and References

The concept of a sliding surface originated in the Soviet literature [e.g., Aizerman and Gantmacher,

1957; Emelyanov, 1957; Filippov, 1960] (see also [Tsypkin, 1955; Flugge-Lotz, et al, 1958]),

mostly in the context of "variable-structure" regulation of linear systems, see [Utkin, 1977] for a

review (also [Young, 1978]). Classical sliding mode control, however, had important drawbacks

limiting its practical applicability, such as chattering and large control authority. The development of

sections 7.1-7.3 is based on [Slotine, 1984].

The combination of sliding controllers with state observers is discussed in [Bondarev, et al.,

1985] in the linear case, and [Hedrick and Ragavan, 1990] in the nonlinear case. Observers based on

sliding surfaces are discussed in [Drakunov, 1983; Slotine, et al, 1986, 1987; Walcott and Zak,

1987]. The development of section 7.4 is adapted from [Slotine, 1985; Hedrick and Gopalswamy,

1989]. The reader is referred to, e.g., [Luenberger, 1979] for a simple proof of the Frobenius-Perron

theorem. Some details on boundary layer interpolations in the multi-input case can be found in

[Slotine, 1985].

Practical implementations of sliding control are described, e.g., in [Yoerger, et al, 1986]

(underwater vehicles), [Hedrick, et al, 1988] (automotive applications), [Harashima, et al, 1988]

(robot manipulators). The literature in the field has been extremely active in the past few years.

Related approaches to robust control include, e.g., [Corless and Leitmann, 1981; Gutman and

Palmor, 1982; Ha and Gilbert, 1985].

7.7 Exercises

7.1 Consider the underwater vehicle of Example 7.4, and assume there is an unmodeled

"pendulum mode" of the vehicle at 2 Hz. Choose X accordingly, and simulate the vehicle's

performance on the same trajectory as in Example 7.4. What is the minimum sampling rate required

to implement your design?

Discuss the performance of the system on various trajectories, which you may want to

generate using a reference model, as in equation (II.5), with ki = 2X,k2 = X2 .

Simulate the unmodeled 2 Hz mode by first passing the control law through a second-order

lowpass filter of unit d.c. gain before actually inputting it in the system. Tune X "experimentally"

around the value given by (7.41), for different values of the filter's damping.

7.2 For the system

Xj = sinx2 + ^ I + I Xj

x2 = a1(()x14cosx2 + a.2(0 "

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (322)

308 Sliding Control Chap. 7

design a controller to track an arbitrary desired trajectory xdt(t) . Assume that the state [jr, x 2 ] r

is measured, that xdl(t) ,xd^(t) ,xdx(t) are all known, and that aft) and a2(f) are unknown time-

varying functions verifying the known bounds

V«>0 , lotjCOl < 10 I < a 2 ( ( ) < 2

Write the full expression of the controller, as a function of the measured state [X\ x2]T • Check

your design in simple simulations. (Hint: First feedback linearize the system by differentiating the

first equation.)

7.3 Consider again the system of Exercise 7.1, but define i so as to use integral control (as

discussed in section 7.1.3). Simulate the controller's performance and compare with the results of

Exercise 7.1, on various trajectories. Show that, besides allowing the system to always start at s = 0,

the integral control term keeps adjusting the control law as long as the effects of bounded

disturbances are not compensated.

7.4 Consider a system of the form (7.1), but where b is now constant and of known constant

positive bounds. Divide both sides of the equation by 6, and write the sliding condition as

where h = 1 Ib . By representing uncertainty on h additively, design a simple switching controller to

satisfy the above condition.

Smooth out the switching controller in boundary layers, and derive the corresponding balance

conditions.

Show that the accuracy of the approximate "bandwidth" analysis in the boundary layer

increases with X.

7.5 Consider a system of the form (7.1), and assume that not only x but also s can be measured

(e.g., that jtW can be measured). An additional term of the form - a s (with a > 0 ) can then be

included in the control law.

Write the corresponding expression of s. Assuming for simplicity that the gain margin p is

constant, show that, given the bandwidth X and the parametric uncertainty, the effect of the

additional term in u is to reduce the maximum value of s by a factor ( 1 + a / p ) . Show that this

implies that the tracking error x can in principle be made arbitrarily small simply by increasing a.

What are the limitations of this approach? In particular, assume that there is / % uncertainty

on the value of x^ . How small must / be to make the approach worthwhile? (Adapted from [Asada

andSlotine, 1986].)

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (323)

Sect. 7.7 Exercises 309

7.6 Show that a condition similar to (7.68) can also be obtained by requiring that the system take

at least two sampling periods to cross the boundary layer.

Assume that sampling is the active limit on bandwidth, and that chattering is acceptable.

Based on the above result, how do the tracking performances of a switching controller and a smooth

sliding controller compare?

7.7 Design a switching controller for the system

x + <X|(/) |JC| Jc2 + a2(t) x3 cos 2x = 5 ii + u

where (X[(/) and a2(r) are unknown time-varying functions verifying the known bounds

V i > 0 , |a,(f) |< 1 - 1 <a 2 ( / )<5

(Hint: let v = 5 u + u . Discuss the effect of chattering in v .)

7.8 In the context of section 7.2, define

sA = s - Osat(.s7<I>)

as a measure of distance to the boundary layer. Show that the time derivative of s^ is well defined

and continuous in time, and, specifically, that one can write

Show that equation (7.26) can be written

5 S ' A 2 * - W

(Adapted from [Slotine and Coetsee, 1986].)

7.9 In the context of tracking control, discuss alternate definitions of 5. For instance, consider

choices based on Butterworth filters rather than (7.3), and how they would modify bounds (7.4).

7.10 Design a sliding controller for the system

x'3 ) + (Xj(r) x- + a2(t) i5 sin 4x = b(t) u

where a.\(t), a2(t), and b(t) are unknown time-varying functions verifying the known bounds

V r > 0 , |ct,(/) |<l |a 2 ( r ) |<2 l<b(t)<4

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (324)

310 Sliding Control Chap. 7

Assume that the state is measured, and that the slowest unmodeled dynamics is the actuator

dynamics, with a time-constant of about 1/50. Simulate the performance of the system on various

trajectories (which you may want to generate using a reference model).

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (325)

Chapter 8Adaptive Control

Many dynamic systems to be controlled have constant or slowly-varying uncertainparameters. For instance, robot manipulators may carry large objects with unknowninertial parameters. Power systems may be subjected to large variations in loadingconditions. Fire-fighting aircraft may experience considerable mass changes as theyload and unload large quantities of water. Adaptive control is a approach to thecontrol of such systems. The basic idea in adaptive control is to estimate the uncertainplant parameters (or, equivalently, the corresponding controller parameters) on-linebased on the measured system signals, and use the estimated parameters in the controlinput computation. An adaptive control system can thus be regarded as a controlsystem with on-line parameter estimation. Since adaptive control systems, whetherdeveloped for linear plants or for nonlinear plants, are inherently nonlinear, theiranalysis and design is intimately connected with the materials presented in this book,and in particular with Lyapunov theory.

Research in adaptive control started in the early 1950's in connection with thedesign of autopilots for high-performance aircraft, which operate at a wide range ofspeeds and altitudes and thus experience large parameter variations. Adaptive controlwas proposed as a way of automatically adjusting the controller parameters in the faceof changing aircraft dynamics. But interest in the subject soon diminished due to thelack of insights and the crash of a test flight. It is only in the last decade that acoherent theory of adaptive control has been developed, using various tools fromnonlinear control theory. These theoretical advances, together with the availability ofcheap computation, have lead to many practical applications, in areas such as robotic

311

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (326)

312 Adaptive Control Chap. 8

: manipulation, aircraft and rocket control, chemical processes, power systems, shipsteering, and bioengineering.

The objective of this chapter is to describe the main techniques and results inadaptive control. We shall start with intuitive concepts, and then study moresystematically adaptive controller design and analysis for linear and nonlinearsystems. The study in this chapter is mainly concerned with single-input systems.Adaptive control designs for some complex multi-input physical systems are discussedin chapter 9.

8.1 Basic Concepts in Adaptive Control

In this section, we address a few basic questions, namely, why we need adaptivecontrol, what the basic structures of adaptive control systems are, and how to go aboutdesigning adaptive control systems.

8.1.1 Why Adaptive Control ?

In some control tasks, such as those in robot manipulation, the systems to becontrolled have parameter uncertainty at the beginning of the control operation. Unlesssuch parameter uncertainty is gradually reduced on-line by an adaptation or estimationmechanism, it may cause inaccuracy or instability for the control systems. In manyother tasks, such as those in power systems, the system dynamics may have wellknown dynamics at the beginning, but experience unpredictable parameter variationsas the control operation goes on. Without continuous "redesign" of the controller, theinitially appropriate controller design may not be able to control the changing plantwell. Generally, the basic objective of adaptive control is to maintain consistentperformance of a system in the presence of uncertainty or unknown variation in plantparameters. Since such parameter uncertainty or variation occurs in many practicalproblems, adaptive control is useful in many industrial contexts. These include:

• Robot manipulation: Robots have to manipulate loads of various sizes, weights,and mass distributions (Figure 8.1). It is very restrictive to assume that the inertialparameters of the loads are well known before a robot picks them up and movesthem away. If controllers with constant gains are used and the load parameters arenot accurately known, robot motion can be either inaccurate or unstable. Adaptivecontrol, on the other hand, allows robots to move loads of unknown parameterswith high speed and high accuracy.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (327)

Sect. 8.1 Basic Concepts in Adaptive Control 313

• Ship steering: On long courses, ships are usually put under automatic steering.However, the dynamic characteristics of a ship strongly depend on many uncertainparameters, such as water depth, ship loading, and wind and wave conditions(Figure 8.2). Adaptive control can be used to achieve good control performanceunder varying operating conditions, as well as to avoid energy loss due to excessiverudder motion.

• Aircraft control: The dynamic behavior of an aircraft depends on its altitude,speed, and configuration. The ratio of variations of some parameters can liebetween 10 to 50 in a given flight. As mentioned earlier, adaptive control wasoriginally developed to achieve consistent aircraft performance over a large flightenvelope.

• Process control: Models for metallurgical and chemical processes are usuallycomplex and also hard to obtain. The parameters characterizing the processes varyfrom batch to batch. Furthermore, the working conditions are usually time-varying{e.g., reactor characteristics vary during the reactor's life, the raw materialsentering the process are never exactly the same, atmospheric and climaticconditions also tend to -change). In fact, process control is one of the mostimportant and active application areas of adaptive control.

desired trajectoryobstacles

•• unknown load

Figure 8.1 : A robot carrying a load of uncertain mass properties

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (328)

314 Adaptive Control Chap. 8

Adaptive control has also been applied to other areas, such as power systemsand biomedical engineering. Most adaptive control applications are aimed at handlinginevitable parameter variation or parameter uncertainty. However, in someapplications, particularly in process control, where hundreds of control loops may bepresent in a given system, adaptive control is also used to reduce the number of designparameters to be manually tuned, thus yielding an increase in engineering efficiencyand practicality.

Figure 8.2 : A freight ship under various loadings and sea conditions

To gain insights about the behavior of the adaptive control systems and also toavoid mathematical difficulties, we shall assume the unknown plant parameters areconstant in analyzing the adaptive control designs. In practice, the adaptive controlsystems are often used to handle time-varying unknown parameters. In order for theanalysis results to be applicable to these practical cases, the time-varying plantparameters must vary considerably slower than the parameter adaptation. Fortunately,this is often satisfied in practice. Note that fast parameter variations may also indicatethat the modeling is inadequate and that the dynamics causing the parameter changesshould be additionally modeled.

Finally, let us note that robust control can also be used to deal with parameteruncertainty, as seen in chapter 7. Thus, one may naturally wonder about thedifferences and relations between the robust approach and the adaptive approach. Inprinciple, adaptive control is superior to robust control in dealing with uncertainties inconstant or slowly-varying parameters. The basic reason lies in the learning behaviorof adaptive control systems: an adaptive controller improves its performance asadaptation goes on, while a robust controller simply attempts to keep consistentperformance. Another reason is that an adaptive controller requires little or no apriori information about the unknown parameters, while a robust controller usuallyrequires reasonable a priori estimates of the parameter bounds. Conversely, robustcontrol has some desirable features which adaptive control does not have, such as itsability to deal with disturbances, quickly varying parameters, and unmodeled

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (329)

Sect. 8.1 Basic Concepts in Adaptive Control 315

dynamics. Such features actually may be combined with adaptive control, leading torobust adaptive controllers in which uncertainties on constant or slowly-varyingparameters is reduced by parameter adaptation and other sources of uncertainty arehandled by robustification techniques. It is also important to point out that existingadaptive techniques for nonlinear systems generally require a linear parametrizationof the plant dynamics, i.e., that parametric uncertainty be expressed linearly in termsof a set of unknown parameters. In some cases, full linear parametrization and thusadaptive control cannot be achieved, but robust control (or adaptive control withrobustifying terms) may be possible.

8.1.2 What Is Adaptive Control ?

An adaptive controller differs from an ordinary controller in that the controllerparameters are variable, and there is a mechanism for adjusting these parameters on-line based on signals in the system. There are two main approaches for constructingadaptive controllers. One is the so-called model-reference adaptive control method,and the other is the so-called self-tuning method.

MODEL-REFERENCE ADAPTIVE CONTROL (MRAC)

Generally, a model-reference adaptive control system can be schematicallyrepresented by Figure 8.3. It is composed of four parts: a plant containing unknownparameters, a reference model for compactly specifying the desired output of thecontrol system, a feedback control law containing adjustable parameters, and anadaptation mechanism for updating the adjustable parameters.

a — estimated parameters

Figure 8.3 : A model-reference adaptive control system

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (330)

316 Adaptive Control Chap. 8

The plant is assumed to have a known structure, although the parameters areunknown. For linear plants, this means that the number of poles and the number ofzeros are assumed to be known, but that the locations of these poles and zeros are not.For nonlinear plants, this implies that the structure of the dynamic equations is known,but that some parameters are not.

A reference model is used to specify the ideal response of the adaptive controlsystem to the external command. Intuitively, it provides the ideal plant responsewhich the adaptation mechanism should seek in adjusting the parameters. The choiceof the reference model is part of the adaptive control system design. This choice has tosatisfy two requirements. On the one hand, it should reflect the performancespecification in the control tasks, such as rise time, settling time, overshoot orfrequency domain characteristics. On the other hand, this ideal behavior should beachievable for the adaptive control system, i.e., there are some inherent constraints onthe structure of the reference model (e.g., its order and relative degree) given theassumed structure of the plant model.

The controller is usually parameterized by a number of adjustable parameters(implying that one may obtain a family of controllers by assigning various values tothe adjustable parameters). The controller should have perfect tracking capacity inorder to allow the possibility of tracking convergence. That is, when the plantparameters are exactly known, the corresponding controller parameters should makethe plant output identical to that of the reference model. When the plant parametersare not known, the adaptation mechanism will adjust the controller parameters so thatperfect tracking is asymptotically achieved. If the control law is linear in terms of theadjustable parameters, it is said to be linearly parameterized. Existing adaptivecontrol designs normally require linear parametrization of the controller in order toobtain adaptation mechanisms with guaranteed stability and tracking convergence.

The adaptation mechanism is used to adjust the parameters in the control law.In MRAC systems, the adaptation law searches for parameters such that the responseof the plant under adaptive control becomes the same as that of the reference model,i.e., the objective of the adaptation is to make the tracking error converge to zero.Clearly, the main difference from conventional control lies in the existence of thismechanism. The main issue in adaptation design is to synthesize an adaptationmechanism which will guarantee that the control system remains stable and thetracking error converges to zero as the parameters are varied. Many formalisms innonlinear control can be used to this end, such as Lyapunov theory, hyperstabilitytheory, and passivity theory. Although the application of one formalism may be moreconvenient than that of another, the results are often equivalent. In this chapter, weshall mostly use Lyapunov theory. J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (331)

Sect. 8.1 Basic Concepts in Adaptive Control 317

As an illustration of MRAC control, let us describe a simple adaptive controlsystem for an unknown mass.

Example 8.1: MRAC control of an unknown mass

Consider the control of a mass on a frictionless surface by a motor force w, with the plant

dynamics being

mx = u (8.1)

Assume that a human operator provides the positioning command r(t) to the control system

(possibly through a joystick). A reasonable way of specifying the ideal response of the controlled

mass to the external command r(t) is to use the following reference model

with the positive constants Xi and "k^ chosen to reflect the performance specifications, and the

reference model output xm being the ideal output of the control system (i.e., ideally, the mass

should go to the specified position r{t) like a well-damped mass-spring-damper system).

If the mass m is known exactly, we can use the following control law to achieve perfect

tracking

u = m(xm — 2Xx — X2x)

with x = x(t) - xm(t) representing the tracking error and X a strictly positive number. This control

law leads to the exponentially convergent tracking error dynamics

x + 2Xx + X2x = 0

Now let us assume that the mass is not known exactly. We may use the following control law

u = m(xm-2Xic-X2x) (8.3)

which contains the adjustable parameter m. Substitution of this control law into the plant

dynamics leads to the closed-loop error dynamics

mv (8.4)

where s, a combined tracking error measure, is defined by

s = i + Xx (8.5)

the signal quantity v by

v = xm -2Xx -X2x

and the parameter estimation error m by

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (332)

318 Adaptive Control Chap. 8

fh = m-m

Equation (8.4) indicates that the combined tracking error s is related to the parameter error

through a stable filter relation.

to use the following update law

through a stable filter relation. One way of adjusting parameter m (for reasons to be seen later) is

m = -yvs (8.6)

where y is a positive constant called the adaptation gain. One easily sees the nonlinear nature of

the adaptive control system, by noting that the parameter m is adjusted based on system signals,

and thus the controller (8.3) is nonlinear.

The stability and convergence of this adaptive control system can be analyzed using

Lyapunov theory. For the closed-loop dynamics (8.4) and (8.6), with s and m as states, we can

consider the following Lyapunov function candidate,

V = -[ms2+-m2] (8.7)2 7

Its derivative can be easily shown to be

V = -Xms2 (8.8)

Using Barbalat's lemma in chapter 4, one can easily show that s converges to zero. Due to the

relation (8.5), the convergence of s to zero implies that of the position tracking error x and the

velocity tracking error x.

For illustration, simulations of this simple adaptive control system are provided in Figures

8.4 and 8.5. The true mass is assumed to be m=2. The initial value of m is chosen to be zero,

indicating no a priori parameter knowledge. The adaptation gain is chosen to be y = 0.5, and the

other design parameters are taken to be X^ = 10, Xj = 25, X. = 6. Figure 8.4 shows the results

when the commanded position is r(t) = 0, with initial conditions being x(0) = xm(Q) — 0 and

x(0) = xm(0) = 0.5. Figure 8.5 shows the results when the desired position is a sinusoidal signal,

r(i) = sin(4 t). It is clear that the position tracking errors in both cases converge to zero, while the

parameter error converge to zero only for the latter case. The reason for the non-convergence of

parameter error in the first case can be explained by the simplicity of the tracking task: the

asymptotic tracking of xm(t) can be achieved by many possible values of the estimated parameter

m, besides the true parameter. Therefore, the parameter adaptation law does not bother to find out

the true parameter. On the other hand, the convergence of the parameter error in Figure 8.5 is

because of the complexity of the tracking task, i.e., tracking error convergence can be achieved

only when the true mass is used in the control law. One may examine Equation (8.4) to see the

mathematical evidence for these statements (also see Exercise 8.1). It is also helpful for the

readers to sketch the specific structure of this adaptive mass control system. A more detailed

discussion of parameter convergence is provided in section 8.2. D

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (333)

Sect. 8.1 Basic Concepts in Adaptive Control 319

Ito

0.6

°-50.4

0-3

0.2

0.1

0.0

-0.1

2.5

UIIJ

<u

ter

para

mel

2.0

1.5

1.0

0.5

0.0 0.5 1.0 1.5 2.0 2.5 3.0time(sec)

0.00.0 0.5 1.0 1.5 2.0 2.5 3.0

time(sec)

Figure 8.4 : Tracking Performance and Parameter Estimation for an Unknown Mass,

0.8

0.6

0.4

0.2

0.0

-0.2

-0.4

-0.6

-0.80.0 0.5 1.0 1.5 2.0 2.5 3.0

time(sec)

0.00.0 0.5 1.5 2.0 2.5 3.0

time(sec)

Figure 8.5 : Tracking Performance and Parameter Estimation for an Unknown Mass,

SELF-TUNING CONTROLLERS (STC)

In non-adaptive control design (e.g., pole placement), one computes the parameters ofthe controllers from those of the plant. If the plant parameters are not known, it isintuitively reasonable to replace them by their estimated values, as provided by aparameter estimator. A controller thus obtained by coupling a controller with an on-line (recursive) parameter estimator is called a self-tuning controller. Figure 8.6illustrates the schematic structure of such an adaptive controller. Thus, a self-tuningcontroller is a controller which performs simultaneous identification of the unknownplant.

The operation of a self-tuning controller is as follows: at each time instant, theestimator sends to the controller a set of estimated plant parameters (a in Figure 8.6),

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (334)

320 Adaptive Control Chap. 8

which is computed based on the past plant input u and output y; the computer finds thecorresponding controller parameters, and then computes a control input u based on thecontroller parameters and measured signals; this control input u causes a new plantoutput to be generated, and the whole cycle of parameter and input updates isrepeated. Note that the controller parameters are computed from the estimates of theplant parameters as if they were the true plant parameters. This idea is often called thecertainty equivalence principle.

controller plant

estimator

S — estimated parameters

Figure 8.6 : A self-tuning controller

Parameter estimation can be understood simply as the process of finding a set ofparameters that fits the available input-output data from a plant. This is different fromparameter adaptation in MRAC systems, where the parameters are adjusted so that thetracking errors converge to zero. For linear plants, many techniques are available toestimate the unknown parameters of the plant. The most popular one is the least-squares method and its extensions. There are also many control techniques for linearplants, such as pole-placement, PID, LQR (linear quadratic control), minimumvariance control, or H°° designs. By coupling different control and estimationschemes, one can obtain a variety of self-tuning regulators. The self-tuning methodcan also be applied to some nonlinear systems without any conceptual difference.

In the basic approach to self-tuning control, one estimates the plant parametersand then computes the controller parameters. Such a scheme is often called indirectadaptive control, because of the need to translate the estimated parameters intocontroller parameters. It is possible to eliminate this part of the computation. To dothis, one notes that the control law parameters and plant parameters are related to eachother for a specific control method. This implies that we may reparameterize the plantmodel using controller parameters (which are also unknown, of course), and then usestandard estimation techniques on such a model. Since no translation is needed in thisscheme, it is called a direct adaptive control scheme. In MRAC systems, one can

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (335)

Sect. 8.1 Basic Concepts in Adaptive Control 321

similarly consider direct and indirect ways of updating the controller parameters.

Example 8.2 : Self-tuning control of the unknown mass

Consider the self-tuning control of the mass of Example 8.1. Let us still use the pole-placement

(placing the poles of the tracking error dynamics) control law (8.3) for generating the control

input, but let us now generate the estimated mass parameter using an estimation law.

Assume, for simplicity, that the acceleration can be measured by an accelerometer. Since the

only unknown variable in Equation (8.1) is m, the simplest way of estimating it is to simply

divide the control input u{t) by the acceleration x , i.e.,

m(t) = ^ (8.9)x

However, this is not a good method because there may be considerable noise in the measurement

x, and, furthermore, the acceleration may be close to zero. A better approach is to estimate the

parameter using a least-squares approach, i.e., choosing the estimate in such a way that the total

prediction error

J=!e2(r)dr (8.10)

is minimal, with the prediction error e defined as

e{t) = m(t)x(t)-u(f)

The prediction error is simply the error in fitting the known input u using the estimated parameter

m. This total error mini:

The resulting estimate is

m. This total error minimization can potentially average out the effects of measurement noise.

wudr•*o

f w2dr

(8.11)

with w = 'x. If, actually, the unknown parameter m is slowly time-varying, the above estimate has

to be recalculated at every new time instant. To increase computational efficiency, it is desirable

to adopt a recursive formulation instead of repeatedly using (8.11). To do this, we define

P(t) = ~ (8.12)

f wzdr* o

The function P(t) is called the estimation gain, and its update can be directly obtained by using

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (336)

322 Adaptive Control Chap. 8

i[P-']=w2 (8.13)at

Then, differentiation of Equation (8.11) (which can be written P ~ ' m= f w u dr ) leads to}o

m = -P(t)we (8.14)

In implementation, the parameter estimate m is obtained by numerically integrating equations

(8.13) and (8.14), instead of using (8.11). Note that a number of other estimation methods can be

used to provide the estimate of the mass. Such methods and their properties are discussed in more

detail in section 8.7, together with a technique for avoiding the use of acceleration measurement

x, using instead velocity or position measurements. Q

It is seen from this example that, in self-tuning control, estimator design andcontroller design are separated. The estimation law (using y and u) is independent ofthe choice of the control law, unlike in MRAC design where the parameter adaptationlaw is affected by the choice of the control law (it is also interesting to note that, inself-tuning control, saturation of the control input has no direct consequence on theconvergence of parameter estimation). While this implies flexibility in design andsimplicity in concept, the analysis of the convergence and stability of the self-tuningcontrol system is usually more complicated.

RELATIONS BETWEEN MRAC AND ST METHODS

As described above, MRAC control and ST control arise from different perspectives,with the parameters in MRAC systems being updated so as to minimize the trackingerrors between the plant output and reference model output, and the parameters in STsystems being updated so as to minimize the data-fitting error in input-outputmeasurements. However, there are strong relations between the two designmethodologies. Comparing Figures 8.3 and 8.6, we note that the two kinds of systemsboth have an inner loop for control and an outer loop for parameter estimation. Froma theoretical point of view, it can actually be shown that MRAC and ST controllerscan be put under a unified framework.

The two methods can be quite different in terms of analysis andimplementation. Compared with MRAC controllers, ST controllers are more flexiblebecause of the possibility of coupling various controllers with various estimators (i.e.,the separation of control and estimation). However, the stability and convergence ofself-tuning controllers are generally quite difficult to guarantee, often requiring thesignals in the system to be sufficiently rich so that the estimated parameters convergeto the true parameters. If the signals are not very rich (for example, if the referencesignal is zero or a constant), the estimated parameters may not be close to the truepamareters, and the stability and convergence of the resulting control system may not

1

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (337)

Sect. 8.1 Basic Concepts in Adaptive Control 323

be guaranteed. In this situation, one must either introduce perturbation signals in theinput, or somehow modify the control law. In MRAC systems, however, the stabilityand tracking error convergence are usually guaranteed regardless of the richness of thesignals.

Historically, the MRAC method was developed from optimal control ofdeterministic servomechanisms, while the ST method evolved from the study ofstochastic regulation problems. MRAC systems have usually been considered incontinuous-time form, and ST regulators in discrete time-form. In recent years,discrete-time version of MRAC controllers and continuous versions of ST controllershave also been developed. In this chapter, we shall mostly focus on MRAC systems incontinuous form. Methods for generating estimated parameters for self-tuning controlare discussed in section 8.7.

8.1.3 How To Design Adaptive Controllers ?

In conventional (non-adaptive) control design, a controller structure {e.g., poleplacement) is chosen first, and the parameters of the controller are then computedbased on the known parameters of the plant. In adaptive control, the major differenceis that the plant parameters are unknown, so that the controller parameters have to beprovided by an adaptation law. As a result, the adaptive control design is moreinvolved, with the additional needs of choosing an adaptation law and proving thestability of the system with adaptation.

The design of an adaptive controller usually involves the following three steps:

• choose a control law containing variable parameters

• choose an adaptation law for adjusting those parameters

• analyze the convergence properties of the resulting control system

These steps are clearly seen in Example 8.1.

When one uses the self-tuning approach for linear systems, the first two stepsare quite straightforward, with inventories of control and adaptation (estimation) lawsavailable. The difficulty lies in the analysis. When one uses MRAC design, theadaptive controller is usually found by trial and error. Sometimes, the three steps arecoordinated by the use of an appropriate Lyapunov function, or using some symbolicconstruction tools such as the passivity formalism. For instance, in designing theadaptive control system of Example 8.1, we actually start from guessing the Lyapunovfunction V (as a representation of total error) in (8.7) and choose the control and

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (338)

324 Adaptive Control Chap. 8 | |

adaptation laws so that V decreases. Generally, the choices of control and adaptationlaws in MRAC can be quite complicated, while the analysis of the convergenceproperties are relatively simple.

Before moving on to the application of the above procedure to adaptive controldesign for specific systems, let us derive a basic lemma which will be very useful inguiding our choice of adaptation laws for MRAC systems.

Lemma 8.1: Consider two signals e and <|> related by the following dynamic equation

(8.15)

where e(t) is a scalar output signal, H(p) is a strictly positive real transfer function, kis an unknown constant with known sign, §(t) is a mxl vector function of time, and\(f) is a measurable mxl vector. If the vector § varies according to

<j>(0 = -sgn(£)yev(0 (8.16)

with y being a positive constant, then e(t) and §(t) are globally bounded. Furthermore,if\ is bounded, then

e(t) -> 0 as t -> oo

Note that while (8.15) involves a mixture of time-domain and frequency-domainnotations (with p being the Laplace variable), its meaning is clear: e(t) is the responseof the linear system of SPR transfer function H(p) to the input [k$T(t)\(t)] (witharbitrary initial conditions). Such hybrid notation is common in the adaptive controlliterature, and later on it will save us the definition of intermediate variables.

In words, the above lemma means that if the input signal depends on the outputin the form (8.16), then the whole system is globally stable (i.e., all its states arebounded). Note that this is a feedback system, shown in Figure 8.7, where the plantdynamics, being SPR, have the unique properties discussed in section 4.6.1.

Proof: Let the state-space representation of (8.15) be

x = Ax+b[*<frrvl (8.17a)

e = c r x (8.17b)

Since H(p) is SPR, it follows from the Kalman-Yakubovich lemma in chapter 4 that given a

symmetric positive definite matrix Q, there exists another symmetric positive definite matrix P

such that

A r P + P A = - Q

J

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (339)

Sect. 8.1 Basic Concepts in Adaptive Control 325

SPR

- sgn(k) y V(t)

Figure 8.7 : A system containing a SPR transfer function

Pb =

Let V be a positive definite function of the form

!A!<|>7> (8.18)

Its time derivative along the trajectories of the system defined by (8.17) and (8.16) is

V =

= -xrQx<0 (8.19)

Therefore, the system defined by (8.15) and (8.16) is globally stable. The equations (8.18) and

(8.19) also imply that e and (|> are globally bounded.

If the signal v(/) is bounded, x is also bounded, as seen from (8.17a). This implies the

uniform continuity of V, since its derivative

is then bounded. Application of Barbalat's lemma of chapter 4 then indicates the asymptotic

convergence of e(l) to zero. CI

It is useful to point out that the system defined by (8.15) and (8.16) not onlyguarantees the boundedness of e and <j>, but also that of the whole state x, as seen from(8.18). Note that the state-space realization in (8.17) can be non-minimal (implyingthe possibility of unobservable or uncontrollable modes) provided that theunobservable and uncontrollable modes be stable, according to the Meyer-Kalman-Yakubovich lemma. Intuitively, this is reasonable because stable hidden modes arenot affected by the choice of $.

In our later MRAC designs, the tracking error between the plant output and

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (340)

326 Adaptive Control Chap. 8

reference model output will often be related to the parameter estimation errors by anequation of the form (8.15). Equation (8.16) thus provides a technique for adjustingthe controller parameters while guaranteeing system stability. Clearly, the tracking-error dynamics in (8.4) satisfy the conditions of Lemma 8.1 and the adaptation law isin the form of (8.16).

8.2 Adaptive Control of First-Order Systems

Let us now discuss the adaptive control of first-order plants using the MRAC method,as an illustration of how to design and analyze an adaptive control system. Thedevelopment can also have practical value in itself, because a number of simplesystems of engineering interest may be represented by a first-order model. Forexample, the braking of an automobile, the discharge of an electronic flash, or theflow of fluid from a tank may be approximately represented by a first-orderdifferential equation

y = -apy + bpu (8.20)

where y is the plant output, u is its input, and ap and bp are constant plant parameters.

PROBLEM SPECIFICATION

In the adaptive control problem, the plant parameters ap and bp are assumed to beunknown. Let the desired performance of the adaptive control system be specified bya first-order reference model

where am and bm are constant parameters, and r(t) is a bounded external referencesignal. The parameter am is required to be strictly positive so that the reference modelis stable, and bm is chosen strictly positive without loss of generality. The referencemodel can be represented by its transfer function M

where

M =

with p being the Laplace variable. Note that M is a SPR function.

The objective of the adaptive control design is to formulate a control law, and

i

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (341)

Sect. 8.2 Adaptive Control of First-Order Systems 327

an adaptation law, such that the resulting model following error y{t) - ym

asymptotically converges to zero. In order to accomplish this, we have to assume thesign of the parameter b to be known. This is a quite mild condition, which is oftensatisfied in practice. For example, for the braking of a car, this assumption amounts tothe simple physical knowledge that braking slows down the car.

CHOICE OF CONTROL LAW

As the first step in the adaptive controller design, let us choose the control law to be

u= ar(t)r + ay(t)y (8.22)

where ar and ay are variable feedback gains. With this control law, the closed-loopdynamics are

y = -(ap-aybp)y + arbpr{t) (8.23)

The reason for the choice of control law in (8.22) is clear: it allows thepossibility of perfect model matching. Indeed, if the plant parameters were known, thefollowing values of control parameters

ar =— a =^— (8.24)bp bp

would lead to the closed-loop dynamics

which is identical to the reference model dynamics, and yields zero tracking error. Inthis case, the first term in (8.22) would result in the right d.c. gain, while the secondterm in the control law (8.22) would achieve the dual objectives of canceling the term(- ay) in (8.20) and imposing the desired pole —amy.

In our adaptive control problem, since a and bp are unknown, the control inputwill achieve these objectives adaptively, i.e., the adaptation law will continuouslysearch for the right gains, based on the tracking error y - ym , so as to make y tend toym asymptotically. The structure of the adaptive controller is illustrated in Figure 8.8.

CHOICE OF ADAPTATION LAW

Let us now choose the adaptation law for the parameters ar and ay. Let

e=y-ym

be the tracking error. The parameter errors are defined as the difference between the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (342)

328 Adaptive Control Chap. 8

Figure 8.8 : A MRAC system for the first-order plant

controller parameter provided by the adaptation law and the ideal parameters, i.e.,

5(0 =.ar

ay

"A *"ar-arA *

ay-ay

(8.25)

The dynamics of tracking error can be found by subtracting (8.23) from (8.21),

b a )y + (b ar - bm)r

= - ame + bp(arr + ayy)

This can be conveniently represented as

e =bD ~ ~ 1 ~

—i—(arr+ayy) = —M(arr

(8.26)

(8.27)

with p denoting the Laplace operator.

Relation (8.27) between the parameter errors and tracking error is in the familiarform given by (8.15). Thus, Lemma 8.1 suggests the following adaptation law

ar = - sgn(bp) yer

a =-sgn(b )yey

(8.28a)

(8.28b)

with y being a positive constant representing the adaptation gain. From (8.28), it isseen that sgn(fe ) determines the direction of the search for the proper controllerparameters.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (343)

Sect. 8.2 Adaptive Control of First-Order Systems 329

TRACKING CONVERGENCE ANALYSIS

With the control law and adaptation law chosen above, we can now analyze thesystem's stability and convergence behavior using Lyapunov theory, or equivalentlyLemma 8.1. Specifically, the Lyapunov function candidate

V(e, *) = I e2 + 1 \bp\(a2 + ay

2) (8.29)

can be easily shown to have the following derivative along system trajectories

V=-a e2

Thus, the adaptive control system is globally stable, i.e., the signals e, ar and a arebounded. Furthermore, the global asymptotic convergence of the tracking error e{t) isguaranteed by Barbalat's lemma, because the boundedness of e, ar and ay implies theboundedness of e (according to (8.26)) and therefore the uniform continuity of V.

It is interesting to wonder why the adaptation law (8.28) leads to tracking errorconvergence. To understand this, let us see intuitively how the control parametersshould be changed. Consider, without loss of generality, the case of a positivesgn(&A Assume that at a particular instant t the tracking error e is negative,indicating that the plant output is too small. From (8.20), the control input u should beincreased in order to increase the plant output. From (8.22), an increase of the controlinput u can be achieved by increasing ar (assuming that r{t) is positive). Thus, theadaptation law, with the variation rate of ar depending on the product of sgn(£>), r ande, is intuitively reasonable. A similar reasoning can be made about ay .

The behavior of the adaptive controller is demonstrated in the followingsimulation example.

Example 8.3: A first-order plant

Consider the control of the unstable plant

y = y + 3«

using the previously designed adaptive controller. The plant parameters a =-l,b = 3 are

assumed to be unknown to the adaptive controller. The reference model is chosen to be

i.e., am = 4, bm = 4. The adaptation gain y is chosen to be equal to 2. The initial values of both

parameters of the controller are chosen to be 0, indicating no a priori knowledge. The initial

conditions of the plant and the model are both zero.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (344)

330 Adaptive Control

Two different reference signals are used in the simulation:

Chap. 8

•ga.DO

5.0 r

3.53.02.52.01.51.00.50.0

0.0

• lit) = 4. It is seen from Figure 8.9 that the tracking error converges to zero but the

parameter error does not.

• r(t) - 4 sin(3 f). It is seen from Figure 8.10 that both the tracking error and

parameter error converge to zero. D

1.0 2.0 3.0 4.0 5.0time(sec)

2.0

1.5

1.0

0.5

0.0

g -0-5°- -1.0

-1.5

-2.00.0 1.0 2.0 3.0 4.0 5.0

time(sec)

Figure 8.9 : Tracking performance and parameter estimation, r(t) = 4

2.0 4.0 6.0 8.0 10.0time(sec)

2.0

1.5

1.0

0.5

0.0

-0.5

-1.0

-1.5

-2.00.0 2.0 4.0 6.0 8.0 10.0

time(sec)

Figure 8.10 : Tracking performance and parameter estimation, r(t) = 4 sin(31)

Note that, in the above adaptive control design, although the stability andconvergence of the adaptive controller is guaranteed for any positive j , am and bm , theperformance of the adaptive controller will depend critically on y. If a small gain ischosen, the adaptation will be slow and the transient tracking error will be large.Conversely, the magnitude of the gain and, accordingly, the performance of theadaptive control system, are limited by the excitation of unmodeled dynamics, becausetoo large an adaptation gain will lead to very oscillatory parameters.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (345)

Sect. 8.2 Adaptive Control of First-Order Systems 331

PARAMETER CONVERGENCE ANALYSIS

In order to gain insights about the behavior of adaptive control system, let usunderstand the convergence of estimated parameters. From the simulation results ofExample 8.3, one notes that the estimated parameters converge to the exact parametervalues for one reference signal but not for the other. This prompts us to speculate arelation between the features of the reference signals and parameter convergence, i.e.,the estimated parameters will not converge to the ideal controller parameters unlessthe reference signal r(t) satisfies certain conditions.

Indeed, such a relation between the features of reference signal andconvergence of estimated parameters can be intuitively understood. In MRACsystems, the objective of the adaptation mechanism is to find out the parameters whichdrive the tracking error y - ym to zero. If the reference signal r(t) is very simple, suchas a zero or a constant, it is possible for many vectors of controller parameters, besidesthe ideal parameter vector, to lead to tracking error convergence. Then, the adaptationlaw will not bother to find out the ideal parameters. Let Q. denote the set composed ofall the parameter vectors which can guarantee tracking error convergence for aparticular reference signal history r(t). Then, depending on the initial conditions, thevector of estimated parameters may converge to any point in the set or wonder aroundin the set instead of converging to the true parameters. However, if the referencesignal r(t) is so complex that only the true parameter vector a* = \a* ay*]T can leadto tracking error convergence, then we shall have parameter convergence.

Let us now find out the exact conditions for parameter convergence. We shalluse simplified arguments to avoid tedious details. Note that the output of the stablefilter in (8.27) converges to zero and that its input is easily shown to be uniformlycontinuous. Thus, arr + ayy must converge to zero. From the adaptation law (8.28)and the tracking error convergence, the rate of the parameter estimates converges tozero. Thus, when time t is large, a is almost constant, and

i.e.,

vT(t) 2 = 0 (8.30)

with

v = [r y]T a = [a,. ayf

Here we have one equation (with time-varying coefficients) and two variables. Theissue of parameter convergence is reduced to the question of what conditions the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (346)

332 Adaptive Control Chap. 8

vector [r(t) y(t)]T should satisfy in order for the equation to have a unique zerosolution.

If r(f) is a constant r0 , then for large t,

with a being the d.c. gain of the reference model. Thus,

[r y] = U «]r0

Equation (8.30) becomes

ar + aa = 0

Clearly, this implies that the estimated parameters, instead of converging to zero,converge to a straight line in parameter space. For Example 8.3, with a = 1, the aboveequation implies that the steady state errors of the two parameters should be of equalmagnitudes but opposite signs. This is obviously confirmed in Figure 8.9.

However, when r{t) is such that the corresponding signal vector v(t) satisfies theso called "persistent excitation" condition, we can show that (8.28) will guaranteeparameter convergence. By persistent excitation of v, we mean that there exist strictlypositive constants ccj and T such that for any t > 0,

' + r a 1 l (8.31)

To show parameter convergence, we note that multiplying (8.30) by v(0 andintegrating the equation for a period of time T, leads to

Ct + T T ~

vv J dra = 0

Condition (8.31) implies that the only solution of this equation is a = 0, i.e., parametererror being zero. Intuitively, the persistent excitation of v(f) implies that the vectorsv(f) corresponding to different times t cannot always be linearly dependent.

The only remaining question is the relation between r(t) and the persistentexcitation of \(t). One can easily show that, in the case of the first order plant, thepersistent excitation of v can be guaranteed, if r(t) contains at least one sinusoidalcomponent.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (347)

Sect. 8.2 Adaptive Control of First-Order Systems 333

EXTENSION TO NONLINEAR PLANTS

The same method of adaptive control design can be used for the nonlinear first-orderplant described by the differential equation

y = -apy-cpf(y) + bpu (8.32)

where / is any known nonlinear function. The nonlinearity in these dynamics ischaracterized by its linear parametrization in terms of the unknown constant c. Insteadof using the control law (8.22), we now use the control law

u = ayy + aff(y) + arr (8.33)

where the second term is introduced with the intention of adaptively canceling thenonlinear term.

Substituting this control law into the dynamics (8.32) and subtracting theresulting equation by (8.21), we obtain the error dynamics

1 ~e = —M( ayy + aff(y) + arr)

kr

where the parameter error 5yis defined as

~ A cp

By choosing the adaptation lawA

ay = -sgn(bp)yey (8.34a)

af= ~sgn(bp)yef (8.34b)

ar - -sgn(bp)yer (8.34c)one can similarly show that the tracking error e converges to zero, and the parametererror remains bounded.

As for parameter convergence, similar arguments as before can reveal theconvergence behavior of the estimated parameters. For constant reference inputr = r0, the estimated parameters converge to the line (with a still being the d.c. gain ofthe reference model)

=0

which is a straight line in the three-dimensional parameter space. In order for the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (348)

334 Adaptive Control Chap. 8

parameters to converge to the ideal values, the signal vector v = [r(t) y(t) f(y)]T

should be persistently exciting, i.e., there exists positive constants ocj and T such thatfor any time f > 0,

,r+r\vTdr> a, I

Generally speaking, for linear systems, the convergent estimation of mparameters require at least m/2 sinusoids in the reference signal r(t), as will bediscussed in more detail in sections 8.3 and 8.4. However, for this nonlinear case,such simple relation may not be valid. Usually, the qualitative relation between r{t)and v(0 is dependent on the particular nonlinear functions/(y). It is unclear how manysinusoids in r(t) are necessary to guarantee the persistent excitation of v(t).

The following example illustrates the behavior of the adaptive system for anonlinear plant.

Example 8.4: simulation of a first-order nonlinear plant

Assume that a nonlinear plant is described by the equation

y- (8.35)

This differs from the unstable plant in Example 8.3 in that a quadratic term is introduced in the

plant dynamics.

Let us use the same reference model, initial parameters, and design parameters as in Example

8.3. For the reference signal r(r) = 4, the results are shown in Figure 8.11. It is seen that the

tracking error converges to zero, but the parameter errors are only bounded. For the reference

signal r(l) = 4sin(3(), the results are shown in Figure 8.12. It is noted that the tracking error and

the parameter errors for the three parameters all converge to zero. LJ

In this example, it is interesting to note two points. The first point is that asingle sinusoidal component in r(t) allows three parameters to be estimated. Thesecond point is that the various signals (including a and y) in this system are muchmore oscillatory than those in Example 8.3. Let us understand why. The basic reasonis provided by the observation that nonlinearity usually generates more frequencies,and thus v(t) may contain more sinusoids than r(t). Specifically, in the above example,with /-(0 = 4sin(31), the signal vector v converges to

v(t) = [r(t) yss(t) fss(t)]

where ysjj) is the steady-state response and fss(t) the corresponding function value,

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (349)

Sect. 8.3 Adaptive Control of Linear Systems With Full State Feedback 335

o

KO

gper

f

a

ack

£3

5.04.54.03.53.02.52.01.51.00.50.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0time(sec)

1.81.61.41.21.00.80.60.40.20.0 >

0.0 0.5 1.0 1.5 2.0 2.5 3.0time(sec)

o03

ew

eter

U1E

J1

0.20.0

-0.2-0.4-0.6

-L0-1.2-1.4-1.6-1.8

0.0 0.5 1.0 1.5 2.0 2.5 3.0time(sec)

0.40.20.0

-0.2-0.4-0.6-0.8-1.0-1.2-1.4-1.6

0.0 0.5 1.0 1.5 2.0 2.5 3.0time(sec)

Figure 8.11 : Adaptive control of a first-order nonlinear system, r(t) = 4

upper left: tracking performance

upper right: parameter ar ; lower left: parameter a ; lower right: parameter as

fss(t) = yss2 = 16A2sin2(3r + (J>) = 8 A2( 1 - cos(6? •

where A and <|) are the magnitude and phase shift of the reference model at co = 3.Thus, the signal vector v(t) contains two sinusoids, with/(jy) containing a sinusoid attwice the original frequency. Intuitively, this component at double frequency is thereason for the convergent estimation of the three parameters and the more oscillatorybehavior of the estimated parameters.

Adaptive Control of Linear Systems With Full StateFeedback

Let us now move on to adaptive control design for more general systems. In thissection, we shall study the adaptive control of linear systems when the full state ismeasurable. We consider the adaptive control of «th-order linear systems in

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (350)

336 Adaptive Control Chap. 8

14.0

3.0

2.0

1.0

0.0

-1.0

-2.0

-3.0

-4.00.0 2.0 4.0 6.0 8.0 10.0

time(sec)

1.8

1.6

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0.00.0 2.0 4.0 6.0 8.0 10.0

time(sec)

B -0.5 -

-1.0 -

-1.5 -

-2.06.0 8.0 10.0

time(sec)

0.5

0.0

-0-5

-1.0

-1.50.0 2.0 4.0 6.0 8.0 10.0

time(sec)

Figure 8.12 : Adaptive control of a first-order nonlinear system, r{t) = 4sin(3t)

upper left: tracking performance

upper right: parameter ar ; lower left: parameter a ; lower right: parameter ay-

companion form:

••••+ ao y = (8.36)

where the state components y,y,... ,y(n~^ are measurable. We assume that thecoefficient vector a = [ an ... aj ao ]T is unknown, but that the sign of an is assumedto be known. An example of such systems is the dynamics of a mass-spring-dampersystem

my + cy + ky = u

where we measure position and velocity (possibly with an optical encoder for positionmeasurement, and a tachometer for velocity measurement, or simply numericallydifferentiate the position signals).

The objective of the control system is to make y closely track the response of astable reference model

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (351)

Sect. 8.3 Adaptive Control of Linear Systems With Full State Feedback 337

< W f l ) + *n-\ym{n-l) + ••••+ ° w * = >it) (8.37)

with r{t) being a bounded reference signal.

CHOICE OF CONTROL LAW

Let us define a signal z(t) as follows

with Pj,..., fin being positive constants chosen such that pn + fin_ipn~l + .... + Po is astable (Hurwitz) polynomial. Adding (- anz(t)) to both sides of (8.36) andrearranging, we can rewrite the plant dynamics as

Let us choose the control law to be

«= anz + an_x^»-l'> + ....+aoy = \r(t)a(t) (8.39)

where \(t) = [z(t) y("~V ....y y]T and

A , . r A A A A n T

a(t) = [an an_l .... a, a j '

denotes the estimated parameter vector. This represents a pole-placement controllerwhich places the poles at positions specified by the coefficients P,-. The tracking errore = y - ym then satisfies the closed-loop dynamics

an[eW + pn_,eO»-l) + .... + %e] = vT(t) a(r) (8.40)

where~ A

a =a - a

CHOICE OF ADAPTATION LAWLet us now choose the parameter adaptation law. To do this, let us rewrite the closed-loop error dynamics (8.40) in state space form,

k = Ax + b[(l/an)vTa] (8.41a)

e = cx (8.41b)

where

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (352)

338 Adaptive Control Chap. 8

00

"Po

10

-Pi

01

-P2

. . 0

1

• • - P n - 1

b =

00

01

c = [ l 0 . . 0 0]

Consider the Lyapunov function candidate

where both F and P are symmetric positive definite constant matrices, and P satisfies

PA + ATP = - Q Q = Q r > 0

for a chosen Q. The derivative V can be computed easily as

V = - xrQx + 2aTvbTPx + 2 a r r - l S

Therefore, the adaptation law

a = -TvbrPx (8.42)

leads to

One can easily show the convergence of x using Barbalat's lemma. Therefore, withthe adaptive controller defined by control law (8.39) and adaptation law (8.42), e andits (n-1) derivatives converge to zero. The parameter convergence condition canagain be shown to be the persistent excitation of the vector v. Note that a similardesign can be made for nonlinear systems in the controllability canonical form, asdiscussed in section 8.5.

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (353)

Sect. 8.4 Adaptive Control of Linear Systems With Output Feedback 339

8.4 Adaptive Control of Linear Systems With OutputFeedback

In this section, we consider the adaptive control of linear systems in the presence ofonly output measurement, rather than full state feedback. Design in this case isconsiderably more complicated than when the full state is available. This partly arisesfrom the need to introduce dynamics in the controller structure, since the output onlyprovides partial information about the system state. To appreciate this need, one cansimply recall that in conventional design (no parameter uncertainty) a controllerobtained by multiplying the state with constant gains (pole placement) can stabilizesystems where all states are measured, while additional observer structures must beused for systems where only outputs are measured.

A linear time-invariant system can be represented by the transfer function

where

... + an_xpn~\ + pn

where kp is called the high-frequency gain. The reason for this term is that the plantfrequency response at high frequency verifies

\W(j<o)\ « —P-

i.e., the high frequency response is essentially determined by kp. The relative degree rof this system is r = n-m. In our adaptive control problem, the coefficients a(-, b:(i = 0, 1,..., n—\;j =0, 1, , m—l) and the high frequency gain kp are all assumed tobe unknown.

The desired performance is assumed to be described by a reference model withtransfer function

Wm(p) = kmZ-p- (8.44)Rm

where Zm and Rm are monic Hurwitz polynomials of degrees nm and mm, and km ispositive. It is well known from linear system theory that the relative degree of the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (354)

340 Adaptive Control Chap. 8

reference model has to be larger than or equal to that of the plant in order to allow thepossibility of perfect tracking. Therefore, in our treatment, we will assume thatnm-mm>n-m.

The objective of the design is to determine a control law, and an associatedadaptation law, so that the plant output y asymptotically approaches ym. Indetermining the control input, the output y is assumed to be measured, but nodifferentiation of the output is allowed, so as to avoid the noise amplificationassociated with numerical differentiation. In achieving this design, we assume thefollowing a priori knowledge about the plant:

• the plant order n is known

• the relative degree n- mis known

• the sign of kp is known

• the plant is minimum-phase

Among the above assumptions, the first and the second imply that the modelstructure of the plant is known. The third is required to provide the direction ofparameter adaptation, similarly to (8.28) in section 8.2. The fourth assumption issomewhat restrictive. It is required because we want to achieve convergent tracking inthe adaptive control design. Adaptive control of non-minimum phase systems is still atopic of active research and will not be treated in this chapter.

In section 8.4.1, we discuss output-feedback adaptive control design for linearplants with relative degree one, i.e., plants having one more pole than zeros. Designfor these systems is relatively straightforward. In section 8.4.2, we discuss output-feedback design for plants with higher relative degree. The design andimplementation of adaptive controllers in this case is more complicated because it isnot possible to use SPR functions as reference models.

8.4.1 Linear Systems With Relative Degree One

When the relative degree is 1, i.e., m = n— 1, the reference model can be chosen to beSPR. This choice proves critical in the development of globally convergent adaptivecontrollers.

CHOICE OF CONTROL LAW

To determine the appropriate control law for the adaptive controller, we must firstknow what control law can achieve perfect tracking when the plant parameters are

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (355)

Sect. 8.4 Adaptive Control of Linear Systems With Output Feedback 341

perfectly known. Many controller structures can be used for this purpose. Thefollowing one, although somewhat peculiar, is particularly convenient for lateradaptation design.

Example 8.5: A controller for perfect tracking

Consider the plant described by

kJp + bn)(8.45)

and the reference model

2P +amlP+am2

r(t)wm(P)

y (t)m

Wp(p)

(8.46)

Figure 8.13 : A model-reference control system for relative degree 1

Let the controller be chosen as shown in Figure 8.13, with the control law being

u = 0CiZ + y + kr (8.47)P + bm

where z = u/(p + bm), i.e., z is the output of a first-order filter with input u, and CX| , p | ,p2 , k are

controller parameters. If we take these parameters to be

P| =_am\-ap\

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (356)

342 Adaptive Control Chap. 8

kKp

one can straightforwardly show that the transfer function from the reference input r to the plant

output y is

P

Therefore, perfect tracking is achieved with this control law, i.e., y(t) = ym{t), V/ > 0.

It is interesting to see why the closed-loop transfer function can become exactly the same as

that of the reference model. To do this, we note that the control input in (8.47) is composed of

three parts. The first part in effect replaces the plant zero by the reference model zero, since the

transfer function from Mj to y (see Figure 8.13) is

The second part places the closed-loop poles at the locations of those of the reference model.

This is seen by noting that the transfer function from u0 to y is (Figure 8.13)

The third part of the control law (km/kp)r obviously replaces kp , the high frequency gain of the

plant, by km. As a result of the above three parts, the closed-loop system has the desired transfer

function. D

The controller structure shown in Figure 8.13 for second-order plants can beextended to any plant with relative degree one. The resulting structure of the controlsystem is shown in Figure 8.14, where k* , 0j* ,02* and 0O* represents controllerparameters which lead to perfect tracking when the plant parameters are known.

The structure of this control system can be described as follows. The block forgenerating the filter signal ci)[ represents an (« - l) tn order dynamics, which can bedescribed by

(»1 = Aooj + hu

where (Oj is an (n— l)xl state vector, A is an («- l)x(n-1) matrix, and h is a constantvector such that (A, h) is controllable. The poles of the matrix A are chosen to be the

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (357)

Sect. 8.4 Adaptive Control of Linear Systems With Output Feedback 343

r(t)

Figure 8.14 : A control system with perfect tracking

same as the roots of the polynomial Zm(p), i.e.,

dct[pl-A] =Zm(p)

1 v-

(8.48)

The block for generating the («— l)xl vector oo2 has the same dynamics but with y asinput, i.e.,

d)2 = Aco2 + hy

It is straightforward to discuss the controller parameters in Figure 8.14. The scalargain k* is defined to be

and is intended to modulate the high-frequency gain of the control system. The vector9j* contains (w-1) parameters which intend to cancel the zeros of the plant. Thevector 02* contains (n-1) parameters which, together with the scalar gain 0O* canmove the poles of the closed-loop control system to the locations of the referencemodel poles. Comparing Figure 8.13 and Figure 8.14 will help the reader becomefamiliar with this structure and the corresponding notations.

As before, the control input in this system is a linear combination of thereference signal r(t), the vector signal (al obtained by filtering the control input u, thesignals 0)2 obtained by filtering the plant output y, and the output itself. The controlinput u can thus be written, in terms of the adjustable parameters and the various

(PDF) Slotine • Li APPLIED NONLINEAR CONTROL - …read.pudn.com/downloads97/ebook/394594/applied nonlinear control...· Slotine • Li APPLIED NONLINEAR CONTROL! i. APPLIED NONLINEAR - DOKUMEN.TIPS (358)

344 Adaptive Control Chap. 8

signals, as

u*{t) = k* r + 8 , * « ) 1 + e 2 * a ) 2 + Q*y (8.49)

Corresponding to this control law and any reference input r(t), the output of the plant

is

(8.50)

since these parameters result in perfect tracking. At this point, one easily sees thereason for assuming the plant to be minimum-phase: this allows the plant zeros to becanceled by the controller poles.

In the adaptive control problem, the plant parameters are unknown, and theideal control parameters described above are also unknown. Instead of (8.49), thecontrol law is chosen to be

u = k{t)r + 91(0<»1 + 62(0(»2 + QoW)1 (8-51)

where k(t), 0](O ,92(0 and Q0(t) are controller parameters to be provided by theadaptation law.

CHOICE OF ADAPTATION LAW

For notational simplicity, let 8 be the 2 / ix l vector containing all the controllerparameters, and oo be the 2« x 1 vector containing the corresponding signals, i.e.,

9(0 = WO 6i(f) 92(0 %(t)]T

(o(t) = [r(t) oo,(O oo2(O y(t)]T

Then the control law (8.51) can be compactly written as

u = QT(t)(o{t) (8.52)

Denoting the ideal value of 6 by 0* and the error between 9(0 and 9 by<|>(0 = 9(f) - 9*, the estimated parameters 9(t) can be represented as

9(0 = 9* + (j)(0

Therefore, the control law (8.52) can also be written as

In order to choose an adaptation law so that the tracking error e converges to zero, we