Linear systems: optimal and robust control

This page intentionally left blank 9217_C000.fm Page i Wednesday, January 3, 2007 10:34 AM Half Title Page Linear S...

Author: Alok K. Sinha

350 downloads 2011 Views 6MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

This page intentionally left blank

9217_C000.fm Page i Wednesday, January 3, 2007 10:34 AM

Half Title Page

Linear Systems Optimal and Robust Control

CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 33487-2742 © 2007 by Taylor & Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group, an Informa business No claim to original U.S. Government works Version Date: 20110614 International Standard Book Number-13: 978-1-4200-0888-3 (eBook - PDF) This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www.copyright. com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com and the CRC Press Web site at http://www.crcpress.com

9217_C000.fm Page iii Wednesday, January 3, 2007 10:34 AM

Title Page

Linear Systems Optimal and Robust Control

Alok Sinha


9217_C000.fm Page v Wednesday, January 3, 2007 10:34 AM

To The Loving Memory of My Mother, Grandparents, and Uncle And My Father

9217_C000.fm Page vii Wednesday, January 3, 2007 10:34 AM

A COMMON MODELING ERROR Prakrieh kriyamanani Gunaih karmani sarvasah Ahamkararavimudhatma Karta ’ham iti manyate

While all kinds of work are done by the modes of nature, he whose soul is bewildered by the self-sense thinks “I am the doer.” (The Bhagavad Gita: An English Translation, by S. Radhakrishnan, George Allen and Unwin, 1971, page 143.)

A ROBUST AND OPTIMAL CONTROL ALGORITHM Karmay eva ’dhikaras te Ma phalesu kadacana Ma karmaphalahetur bhur Ma te sango ’stv akarmani

To action alone hast thou a right and never at all to its fruits; let not the fruits of action be thy motive; neither let there be in thee any attachment to inaction. (The Bhagavad Gita: An English Translation, by S. Radhakrishnan, George Allen and Unwin, 1971, page 119.)


9217_bookTOC.fm Page ix Wednesday, January 3, 2007 10:52 AM

Contents Chapter 1 1.1 1.2

Overview ..........................................................................................................1 Contents of the Book .......................................................................................2

Chapter 2 2.1 2.2

Introduction ..........................................................................................1

State Space Description of a Linear System.......................................3

Transfer Function of a Single Input/Single Output (SISO) System ...............3 State Space Realizations of a SISO System....................................................5 Method I...........................................................................................................5 Method II..........................................................................................................7 Properties of State Space Models..................................................................10 1. Duality.................................................................................10 2. Nonuniqueness of State Space Realization........................10 2.3 SISO Transfer Function from a State Space Realization..............................11 2.4 Solution of State Space Equations ................................................................12 2.4.1 Homogeneous Equation .....................................................................12 2.4.2 Inhomogeneous Equation...................................................................13 2.5 Observability and Controllability of a SISO System....................................14 2.5.1 Observability ......................................................................................14 Observability of State Space Realization Using Method I ...............16 Observability of State Space Realization Using Method II..............16 2.5.2 Controllability ....................................................................................19 Controllability of the State Space Realization Obtained Using Method I ..................................................................................24 Controllability of the State Space Realization Obtained Using Method II.................................................................................25 2.6 Some Important Similarity Transformations .................................................29 2.6.1 Diagonal Form ...................................................................................29 2.6.2 Controllability Canonical Form .........................................................30 2.7 Simultaneous Controllability and Observability ...........................................31 2.7.1 Observability of State Space Realization Using Method I ...............32 2.8 Multiinput/Multioutput (MIMO) Systems.....................................................36 2.9 State Space Realizations of a Transfer Function Matrix ..............................40 Method I.........................................................................................................40 Method II........................................................................................................42 2.10 Controllability and Observability of a MIMO System .................................44 2.10.1 Controllability and Observability of Methods I and II Realizations ........................................................................................45 Method I Realization .........................................................................45

9217_bookTOC.fm Page x Wednesday, January 3, 2007 10:52 AM

Method II Realization ........................................................................46 2.11 Matrix-Fraction Description (MFD)..............................................................46 2.11.1 Degree of a Square Polynomial Matrix and Greatest Common Right Divisor (gcrd)...........................................................47 2.11.2 Elementary Row and Column Operations.........................................50 2.11.3 Determination of a gcrd.....................................................................51 2.12 MFD of a Transfer Function Matrix for the Minimal Order of a State Space Realization ...........................................................................................57 2.13 Controller Form Realization from a Right MFD ..........................................59 2.13.1 State Space Realization......................................................................59 2.13.2 Similarity Transformation to Convert Any State Space Realization {A,B,C} to the Controller Form Realization..................64 2.14 Poles and Zeros of a MIMO Transfer Function Matrix ...............................66 2.14.1 Smith Form ........................................................................................66 2.14.2 Smith–McMillan Form ......................................................................67 2.14.3 Poles and Zeros via Smith–McMillan Form .....................................69 2.14.4 Poles and Zeros via an Irreducible MFD ..........................................71 2.15 Stability Analysis ...........................................................................................71 An Important Property of the Lyapunov Equation ...........................72 Exercise Problems ...................................................................................................73 Chapter 3 3.1

3.2 3.3 3.4

3.5

State Feedback Control and Optimization ........................................79

State Variable Feedback for a Single Input System ......................................79 3.1.1 Effects of State Feedback on Poles of the Closed-Loop Transfer Function: Computation of State Feedback Gain Vector .................................................................................................80 3.1.2 Effects of State Feedback on Zeros of the Closed-Loop Transfer Function ...............................................................................85 3.1.3 State Feedback Control for a Nonzero and Constant Output ...........89 3.1.4 State Feedback Control under Constant Input Disturbances: Integral Action....................................................................................91 Computation of State Feedback Gain Matrix for a Multiinput System .......95 State Feedback Gain Matrix for a Multiinput System for Desired Eigenvalues and Eigenvectors........................................................................99 Fundamentals of Optimal Control Theory ..................................................112 3.4.1 Necessary Conditions for Optimality ..............................................113 Minimization of I .............................................................................114 Maximization of I ............................................................................115 3.4.2 Properties of Hamiltonian for an Autonomous System ..................115 Special Case .....................................................................................116 Linear Quadratic Regulator (LQR) Problem ...............................................123 3.5.1 Solution (Open-Loop Optimal Control) ..........................................124 3.5.2 Solution (Closed-Loop Optimal Control)........................................126 3.5.3 Cross Term in the Objective Function.............................................127 3.5.4 Important Case: Infinite Final Time ................................................132

9217_bookTOC.fm Page xi Wednesday, January 3, 2007 10:52 AM

3.6

Solution of LQR Problem Via Root Locus Plot: SISO Case .....................137 The Symmetric Root Locus.........................................................................139 Case I: High Cost of Control (r → ∞)............................................140 Case II: Low Cost of Control (r → 0) ............................................140 3.7 Linear Quadratic Trajectory Control ...........................................................143 3.8 Frequency-Shaped LQ Control ....................................................................148 3.9 Minimum-Time Control of a Linear Time-Invariant System .....................154 Normality of Linear Systems.......................................................................155 Existence and Uniqueness Theorems on Minimum-Time Control.............156 Exercise Problems .................................................................................................168 Chapter 4

Control with Estimated States.......................................................... 179

4.1 4.2

Open-Loop Observer....................................................................................179 Closed-Loop Observer .................................................................................180 4.2.1 Determination of Observer Gain Vector L for a Single-Output System ......................................................................181 4.2.2 Determination of Observer Gain Matrix L for a Multiple-Output System...................................................................182 4.2.3 Locations of Observer Poles............................................................182 4.3 Combined Observer–CONTROLLER .........................................................186 4.4 Reduced-Order Observer .............................................................................191 4.5 Response of a Linear Continuous-Time System to White Noise ...............195 4.6 Kalman Filter: Optimal State Estimation ....................................................200 State Dynamics ............................................................................................200 Measurement Equation ................................................................................200 Observer Dynamics......................................................................................201 4.7 Stochastic Optimal Regulator in Steady State.............................................208 State Dynamics ............................................................................................208 Objective Function .......................................................................................209 4.8 Linear Quadratic Gaussian (LQG) Control .................................................215 State Dynamics ............................................................................................216 Measurement Equation ................................................................................216 Objective Function .......................................................................................216 Closed-Loop System Dynamics...................................................................218 Minimum Value of the Objective Function .................................................219 4.9 Impact of Modeling Errors on Observer-Based Control.............................222 4.9.1 Structured Parametric Uncertainties ................................................222 4.9.2 Unmodeled Dynamics......................................................................224 Exercise Problems .................................................................................................228 Chapter 5

5.1

Robust Control: Fundamental Concepts and H2, H∞, and μ Techniques ....................................................................................235

Important Aspects of Singular Value Analysis ............................................235

9217_bookTOC.fm Page xii Wednesday, January 3, 2007 10:52 AM

5.1.1

Significance of Minimum and Maximum Singular Values of the Transfer Function Matrix at a Frequency .............................235 5.1.2 An Inequality for Robustness Test ..................................................237 5.2 Robustness: Sensitivity and Complementary Sensitivity ............................238 5.2.1 Basic Definitions ..............................................................................238 5.2.2 Robustness to Structured Uncertainties ...........................................242 5.2.3 Robustness to Unstructured Uncertainties.......................................244 5.3. Robustness of LQR and Kalman Filter (KF) Feedback Loops ..................252 5.3.1 Liner Quadratic Regulator (LQR) Feedback Loop .........................252 5.3.2 Gain and Phase Margins of a Single-Input System ........................254 5.3.3 Gain and Phase Margins of a Multiinput System ...........................257 5.3.2 Kalman Filter (KF) Loop.................................................................260 5.4 LQG/LTR Control ........................................................................................263 5.4.1 Lack of Guaranteed Robustness of LQG Control...........................263 5.4.2 Loop Transfer Recovery (LTR) .......................................................265 5.5 H2 And H∞ Norms........................................................................................269 5.5.1 Definition of L2 Norm......................................................................269 5.5.2 Definition of H2 Norm .....................................................................269 5.5.3 Significance of H2 Norm..................................................................270 I. Connection with Unit Impulse Responses .......................270 II. Connection with the Root Mean Square (RMS) Response ...........................................................................271 5.5.4 Definition of H∞ Norm.....................................................................273 5.5.5 Significance of H∞ Norm .................................................................273 I. Connection with the Worst-Case Frequency Response ...........................................................................273 II. Connection with the Worst-Case Output or Input ...........274 III. Connection of the Bounded Real Lemma (5.5.28 to 5.5.30) with a Special Linear Quadratic Maximization Problem......................................................275 5.5.6 Computation of H2 Norm ................................................................278 5.5.7 Computation of H∞ Norm ................................................................281 5.6 H2 Control ....................................................................................................283 5.6.1 Full State Feedback H2 Control (Figure 5.6.1) ...............................283 5.6.2 Output Feedback H2 Control (Figure 5.6.2)....................................286 5.7 Well-Posedness, Internal Stability, and Small Gain Theorem ....................288 5.7.1 Well-Posedness and Internal Stability of a General Feedback System ..............................................................................................288 Well-Posedness of Feedback System ..............................................290 Internal Stability of the System .......................................................290 5.7.2 Small Gain Theorem ........................................................................292 5.7.3 Analysis for Application of Small Gain Theorem ..........................297 5.8 Formulation of Some Robust Control Problems with Unstructured Uncertainties.................................................................................................299 5.8.1 Multiplicative Uncertainty ...............................................................299 5.8.2 Additive Uncertainty ........................................................................301

9217_bookTOC.fm Page xiii Wednesday, January 3, 2007 10:52 AM

5.9

Formulation of Robust Control Problems with Structured Uncertainties.................................................................................................304 5.9.1 Linear Fractional Transformation (LFT) .........................................304 Lower LFT .......................................................................................305 Upper LFT........................................................................................306 5.9.2 Structured Parametric Uncertainties ................................................308 5.10 H∞ Control....................................................................................................314 5.10.1 Full State Feedback H∞ Control (Figure 5.10.1).............................314 5.10.2 Full State Feedback H∞ Control under Disturbance Feedforward (Figure 5.10.3) ............................................................320 5.10.3 Guaranteed H∞ Norm via State Estimation .....................................322 5.10.4 Output Feedback H∞ Control...........................................................324 Equivalence of Stability of Systems Shown in Figure 5.10.6 and Figure 5.10.7 .............................................................................328 Controller Design.............................................................................332 5.10.5 Poles or Zeros on Imaginary Axis 336 5.11 Loop Shaping ...............................................................................................337 5.11.1 Trade-Off between Performance and Robustness via H∞ Control..............................................................................................337 5.11.2 Fundamental Constraint: Bode’s Sensitivity Integrals ....................345 SISO System ....................................................................................345 5.12 Controller Based on μ Analysis...................................................................351 5.12.1 Definitions and Properties of Structured Singular Values (μ) ........351 5.12.2 Robustness Analysis via μ ...............................................................353 5.12.3 Robust Performance via μ Analysis ................................................354 5.12.4 μ Synthesis: D–K Iteration ..............................................................366 Exercise Problems .................................................................................................367 Chapter 6 6.1 6.2

Robust Control: Sliding Mode Methods..........................................377

Basic Concepts of Sliding Modes ...............................................................377 Sliding Mode Control of a Linear System with Full State Feedback ........379 6.2.1 Computation of Sliding Hyperplane Matrix G ...............................380 6.2.2 Optimal Sliding Mode (OS) Controller...........................................383 6.3 Sliding Mode Control of an Uncertain Linear System with Full State Feedback: Blending H∞ and Sliding Mode Methods ..................................386 6.3.1 Impact of Uncertainties on System Dynamics in Sliding Modes ...............................................................................................387 6.3.2 Computation of Sliding Hyperplane Matrix via H∞ Control Method .............................................................................................388 6.4 Sliding Mode Control of a Linear System with Estimated States..............394 6.5 Optimal Sliding Mode Gaussian (OSG) Control ........................................396 Exercise Problems .................................................................................................402

9217_bookTOC.fm Page xiv Wednesday, January 3, 2007 10:52 AM

References .............................................................................................................405 Appendix A: Linear Algebraic Equations, Eigenvalues/Eigenvectors, and Matrix Inversion Lemma ............................................................................409 A.1 System of Linear Algebraic Equations........................................................409 Definitions ....................................................................................................409 Linear Independence of Vectors ......................................................409 Range or Column Space ..................................................................409 Rank of a Matrix..............................................................................409 Null Space ........................................................................................410 A Systematic Test for Linear Independence of Vectors..............................414 A.2 Eigenvalues and Eigenvectors......................................................................415 A.3 Matrix Inversion Lemma .............................................................................415 Appendix B: Quadratic Functions, Important Derivatives, Fourier Integrals, and Parseval’s Relation .....................................................................417 B.1 Quadratic Functions .....................................................................................417 B.2 Derivative of a Quadratic Function .............................................................419 B.3 Derivative of a Linear Function...................................................................421 B.4 Fourier Integrals and Parseval’s Theorem ...................................................422 B.4.1 Scalar Signal ....................................................................................422 Scalar Parseval Relation...................................................................424 B.4.2 Vector Signal ....................................................................................424 Multivariable Parseval Relation .......................................................425 Appendix C: Norms, Singular Values, Supremum, and Infinimum ..............427 C.1 Vector Norms ...............................................................................................427 C.2 Matrix Norms...............................................................................................427 C.3 Singular Values of a Matrix.........................................................................429 C.4 Singular Value Decomposition (SVD).........................................................429 Computation of U and V ..............................................................................430 Significance of Singular Vectors Corresponding to the Maximum and Minimum Singular Values ...........................................................................431 C.5 Properties of Singular Values.......................................................................432 C.6 Supremum and Infinimum ...........................................................................432 Appendix D: Stochastic Processes .....................................................................435 D.1 Stationary Stochastic Process ......................................................................435 Autocorrelation Function .............................................................................436 Autocovariance Function .............................................................................437 Crosscorrelation Function ............................................................................437 Crosscovariance Function ............................................................................437 Uncorrelated Stochastic Processes...................................................438 D.2 Power Spectrum or Power Spectral Density (PSD) ....................................438

9217_bookTOC.fm Page xv Wednesday, January 3, 2007 10:52 AM

D.3 White Noise: A Special Stationary Stochastic Process...............................439 Generation of Approximate Gaussian White Noise in MATLAB ..............439 D.4 Response of a SISO Linear and Time-Invariant System Subjected to a Stationary Stochastic Process ...............................................................440 D.5 Vector Stationary Stochastic Processes .......................................................440 Appendix E: Optimization of a Scalar Function with Constraints in the Form of a Symmetric Real Matrix Equal to Zero ...............................443 Appendix F: A Flexible Tetrahedral Truss Structure ......................................447 Appendix G: Space Shuttle Dynamics during Reentry...................................451 State Space Model .......................................................................................453 Index......................................................................................................................457

9217_book.fm Page 1 Tuesday, December 19, 2006 10:44 AM

1

Introduction

1.1 OVERVIEW This book contains materials on linear systems, and optimal and robust control, and is an outgrowth of two graduate level courses I have taught for many years at The Pennsylvania State University, University Park. The first course is on linear systems and optimal control, whereas the second course is on robust control. The unique feature of this book is that it presents the materials in a theoretically rigorous way while keeping the applications to practical problems in mind. Also, this is the first book containing H∞ and sliding mode methods together. The materials on linear systems include controllability, observability, and matrix fraction description. First, the concepts of state feedback control and observers are developed. Next, the optimal control is presented along with stochastic optimal control. Then, the lack of robustness of LQG control is discussed. This is followed by the presentation of robust control techniques. The derivation of H∞ control theory is developed from the first principle. The sliding mode control of a linear system is presented. Then, it is shown how a blend of sliding mode control and H∞ methods can enhance the robustness of the system. One of the objectives is to make the book self contained as much as possible. For example, all the required concepts for stochastic processes are presented so that a student can understand LQG control without much prior background in stochastic processes. The book contains the presentation of theory with practical examples to illustrate the key theoretical concepts and to show their applications to practical problems. At the end of each chapter, exercise problems are included. The use of MATLAB software has been highlighted. For my course on linear systems and optimal control, I have used the textbook by T. Kailath, Linear Systems (Prentice-Hall, 1980). This is a great book, and contains extensive amount of information on linear systems. For my course on robust control, I have used the textbook by K. Zhou and J. C. Doyle, Essentials of Robust Control (Prentice-Hall, 1998). This is also an excellent book. However, a typical engineering graduate student in most universities may find the materials in both these textbooks to be highly mathematical and, as a result, find them difficult to follow. Therefore, I have developed mathematical analyses in this book by keeping in view the background of a typical engineering student with a bachelor’s degree. For example, the derivation of H∞ does not require students to learn additional mathematical tools. I have learned the linear system theory from the textbook of T. Kailath. Therefore, even though I have never met him, I would like to recognize T. Kailath as my virtual

1


2

Linear Systems: Optimal and Robust Control

teacher. During my sabbatical at MIT, I was fortunate to interact with M. Athans and attend his course on multivariable control systems, for which G. Stein presented excellent lectures on H∞ control. I would also like to thank E. F. Crawley for arranging my sabbatical and providing me an opportunity to do research on sliding mode control at the MIT Space Engineering Research Center, where I was lucky to find David Miller who helped me implement my controller on the development model of the Middeck Active Control Experiment (MACE). Lastly, I would like to thank my wife, Hansa, and daughters, Divya and Swarna, for their support.

1.2 CONTENTS OF THE BOOK In Chapter 2, methods to develop a state space realization from a SISO transfer function are presented, along with the concepts of controllability and observability. The connection between a minimal order of the SISO state space realization and simultaneous controllability and observability is presented. This is followed by the matrix fraction description of the MIMO system, and a method is developed to find a state space realization with the minimal order. Lastly, poles and zeros of a MIMO system are defined. In Chapter 3, the design of a full state feedback control system is presented for a SISO system along with its impact on poles and zeros of the closed-loop system. Next, the full state feedback control system is presented for a MIMO system. The necessary conditions for the optimal control are then derived and used to develop the linear quadratic (LQ) control theory and the minimum time control. In Chapter 4, methods to estimate states are developed on the basis of inputs and outputs of a deterministic and a stochastic system. Then, theories and examples of optimal state estimation and linear quadratic Gaussian control are presented. In Chapter 5, the fundamental concepts of robust control are developed. The robustness of LQ and LQG control techniques developed in Chapter 3 and Chapter 4 are examined. Lastly, theories for H2, H∞, and μ techniques are presented along with Bode’s sensitivity integrals and illustrative examples. In Chapter 6, basic concepts of sliding modes are presented along with the sliding mode control of a linear system with full state feedback. Then, it is shown how H∞ and sliding mode theories can be blended to control an uncertain linear system with full state feedback. Next, the sliding mode control of a deterministic linear system is developed with the feedback of estimated states. Lastly, the optimal sliding Gaussian (OSG) control theory is presented for a stochastic system.


2

State Space Description of a Linear System

First, methods to develop a state space realization from a SISO transfer function are presented, as well as the concepts of controllability and observability. The connection between a minimal order of the SISO state space realization and the simultaneous controllability and observability is presented. This is followed by the matrix fraction description (MFD) of the MIMO system, and a method is developed to find a state space realization with the minimal order. Lastly, poles and zeros of a MIMO system are defined.

2.1 TRANSFER FUNCTION OF A SINGLE INPUT/SINGLE OUTPUT (SISO) SYSTEM The dynamics of a single input/single output (SISO) linear system can be represented in general by the following nth order differential equation: dny d n−1y d n−2 y a a + + + … + an y = 1 2 dt n dt n−1 dt n−2 d n−1u d n−2 u d nu b0 n + b1 n−1 + b2 n−2 + … + bnu dt dt dt

(2.1.1)

where y (t ) is the output and u (t ) is the input. The coefficients a1 , a2 ,…, an and b0 , b1 ,…, bn are system parameters. These parameters are constants for a time-invariant system. Taking the Laplace transformation of (2.1.1) and setting all initial conditions to be zero, y (s ) = g (s ) u (s )

(2.1.2)

where g( s ) =

b0 s n + b1s n−1 + … + bn−1s + bn s + a1s n−1 + a2 s n−2 + …an−1s + an n

(2.1.3)

3


4


y1 ( s )

u(s )

y (s )

g1 ( s)

g2 (s)

FIGURE 2.1 Cascaded linear systems.

The expression g (s ) is described as the transfer function with which the following facts can be attributed: 1. The transfer function is the ratio of Laplace transforms of the output and the input. 2. All initial conditions associated with the input and the output are taken to be zero. 3. The transfer function is only defined for a linear and time-invariant system. Taking u (t ) = δ(t ) , the unit impulse function, it can be seen that g (t ) = L−1 ( g (s ))

(2.1.4)

is the unit impulse response function of the system. In other words, the transfer function of a linear time-invariant system is the Laplace transformation of the unit impulse response of the system. To appreciate the usefulness of the transfer function approach, consider the system shown in Figure 2.1 in which the output of the first subsystem y1 (s ) is the input to the next subsystem. This is a typical situation found in the study or design of a control system. The output y (s ) and the input u (s ) are related as follows: y (s ) = g1 (s ) g2 (s ) u (s )

(2.1.5)

In time domain, the output y (t ) is related to the input u(t) via the convolution integral (Kuo, 1995). More specifically, y (t ) =

t

∫ g (t − τ)y (τ)d τ 2

1

(2.1.6a)

0

and y1 (t ) =

t

∫ g (t − τ)u(τ)d τ 1

0

Therefore,

(2.1.6b)



y (t ) =

5

τ

t

∫ g (t − τ) ∫ g (τ − ν)u(ν)d νd τ 2

0

1

(2.1.7)

0

Comparing Equation 2.1.5 and Equation 2.1.7, it is obvious that the input–output relationship in s-domain is much simpler than that in time domain.

2.2 STATE SPACE REALIZATIONS OF A SISO SYSTEM METHOD I Define ξ(t ) such that d nξ d n−1ξ d n−2 ξ + a1 n−1 + a2 n−2 + … + an ξ = u (t ) n dt dt dt

(2.2.1)

Assuming that all initial conditions on y (t ) are zero and using the principle of superposition (Kailath, 1980), Equation 2.1.1 leads to

y (t ) = b0

d n −1ξ d n −2 ξ d nξ + b1 n −1 + b2 n −2 + ...... + bn ξ n dt dt dt

(2.2.2)

Although initial conditions on y(t) and its higher derivatives have been taken to be zero, the Equation 2.2.2 is valid for nonzero initial conditions on y(t) and its higher 0 ) ,…, and derivatives. The treatment of nonzero y(0 ) , y(

dny (0 ) is related to the dt n

observability issue and will be discussed later. Define x1 = ξ

x2 =

dξ dt

x3 =

d 2ξ dt 2

xn =

d n−1ξ dt n−1

(2.2.3)


6


Using (2.2.1) and (2.2.3), the following n first-order differential equations are obtained as follows: dx1 = x2 dt dx2 = x3 dt dxn−1 = xn dt dxn = −an x1 − an−1x2 − … − a1xn + u (t ) dt

(2.2.4)

Using Equation 2.2.2, the output y (t ) is related to variables defined in (2.2.3) as follows: y(t ) = b0

dxn + b1xn + b2 xn−1 + … + bn x1 dt

(2.2.5)

Using (2.2.4), y (t ) = (bn − b0 an ) x1 + (bn −1 − b0 an −1 ) x2 + ..... + (b1 − b0 a1 ) x n + b0 u (t ) (2.2.6) The system represented by (2.2.4) and (2.2.6) can be realized using n analog integrators as shown in Figure 2.2. The variables x1, x2 , x3,…, x n turn out to be outputs of integrators and are described as state variables. In matrix form, Equation 2.2.4 and Equation 2.2.6 are described as follows: dx = Ac x (t ) + b c u (t ) dt

(2.2.7)

y (t ) = cc x (t ) + b0 u (t )

(2.2.8)

and

where



7

y

b0 b1

bn b2

u

xn

xn

bn xn

1

x2

1

x1

a1 a2

an

1

an

FIGURE 2.2 Analog computer simulation diagram (Method I).

⎡ 0 ⎢ 0 Ac = ⎢ ⎢ . ⎢ ⎢⎣ − an

1 0 . .

0 1 . .

0 . . .

. . . .

. . . .

. . . − a2

⎡0 ⎤ ⎢ ⎥ ⎢0 ⎥ ⎢.⎥ bc = ⎢ ⎥ ⎢.⎥ ⎢.⎥ ⎢ ⎥ ⎢⎣ 1 ⎥⎦

0 ⎤ ⎥ 0 ⎥ . ⎥ ⎥ − a1 ⎥⎦

(2.2.9)

(2.2.10)

and c c = [bn − b0 an

bn−1 − b0 an−1 … b1 − b0 a1 ]

METHOD II Defining p =

d , Equation 2.1.1 can be written as dt

(2.2.11)


8

Linear Systems: Optimal and Robust Control u

bn

bn

x1

x1

an

b1

1

x2

an

x2

xn

b0

xn

1

xn

y

a1

1

FIGURE 2.3 Analog computer simulation diagram (Method II).

p n ( y(t ) − b0 u (t )) + p n−1 (a1 y(t ) − b1u (t )) + p n−2 (a2 y(t ) − b2 u (t )) + … + (an y(t ) − bn u (t )) = 0

(2.2.12)

Dividing (2.2.12) by p n (Wiberg, 1971), y(t ) = b0 u (t ) −

a1 y(t ) − b1u (t ) a2 y(t ) − b2 u (t ) a y(t ) − bn u (t ) (2.2.13) − −… − n 2 p p pn

The analog computer simulation diagram corresponding to Equation 2.2.13 is shown in Figure 2.3. Defining the outputs of integrators as state variables x1, x2 , x3,…, x n , the following relationships are obtained: y (t ) = x n (t ) + b0 u (t ) and dx1 = − an y (t ) + bn u (t ) = − an x n + (bn − b0 an )u dt dx2 = −an−1y(t ) + x1 (t ) + bn−1u (t ) = −an−1xn + x 1+(bbn−1 − b0 an−1 )u dt dxn = −a1y(t ) + xn−1 (t ) + b1u (t ) = −a1xn + x n−1+(b1 − b0 a1 )u dt

(2.2.14)



9

Equations 2.2.14 can be put in the matrix form as follows: dx = Ao x (t ) + b o u (t ) dt

(2.2.15a)

y (t ) = co x (t ) + b0 u (t )

(2.2.15b)

and

where ⎡0 ⎢ ⎢1 Ao = ⎢0 ⎢ ⎢. ⎢0 ⎣

0 0 1 . 0

. . . . .

. . . . .

− an ⎤ ⎥ − an −1 ⎥ − an −2 ⎥ ⎥ . ⎥ − a1 ⎥⎦

. 0 0 . 1

⎡ bn − b0 an ⎤ ⎢ ⎥ ⎢bn −1 − b0 an −1 ⎥ ⎢ ⎥ . bo = ⎢ ⎥ . ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ ⎢⎣ b1 − bo a1 ⎥⎦

(2.2.16)

(2.2.17)

and co = ⎡⎣0

0

.

.

0

1⎤⎦

(2.2.18)

Notation Often, a state space realization is represented by the symbol {A, b, c} by which we mean the following: x = Ax (t ) + bu (t )

(2.2.19a)

y (t ) = cx (t )

(2.2.19b)


10


EXAMPLE 2.1

g (s ) =

s 3 + 2s + 5 s + 2 s 2 + 4s + 3

(2.2.20)

3

Here, b0 = 1. From Method I: ⎡0 ⎢ Ac = ⎢ 0 ⎢ −3 ⎣

1 0 −4

0⎤ ⎥ 1⎥ −2 ⎥⎦

⎡0 ⎤ ⎢ ⎥ b c = ⎢0 ⎥ ⎢1⎥ ⎣ ⎦

cc = ⎡⎣2

2

2 ⎤⎦

(2.2.21)

From Method II: ⎡0 ⎢ Ao = ⎢ 1 ⎢0 ⎣

PROPERTIES 1.

OF

−3 ⎤ ⎥ −4 ⎥ −2 ⎥⎦

0 0 1

⎡2 ⎤ ⎢ ⎥ b o = ⎢2 ⎥ ⎢2 ⎥ ⎣ ⎦

co = ⎡⎣0

0

1⎤⎦

(2.2.22)

STATE SPACE MODELS

Duality

{

}

For any given realization A, b, c of a system transfer function, there is a dual (Kailath, 1980) realization AT , cT , b T . It is interesting to note that

{

}

Ao = Ac T, b o = ccT , and co = b cT

(2.2.23)

Hence, the state space realization obtained by Method I is dual to that obtained by Method II, and vice versa. 2.

Nonuniqueness of State Space Realization

Consider the following transformation: x (t ) = Tx (t )

(2.2.24)

where T is any nonsingular matrix. For the realization {A, b, c}, dx = Ax + bu (t ) dt

(2.2.25)



11

y (t ) = cx (t )

(2.2.26)

Substituting (2.2.24) into (2.2.25) and (2.2.26), dx = Ax + bu (t ) dt

(2.2.27)

y (t ) = cx (t )

(2.2.28)

where A = T −1 AT ,

b = T −1b ,

and

c = cT

(2.2.29)

Hence, a new realization, {A, b, c }, has been obtained. As the choice of the nonsingular matrix T is arbitrary, there are clearly many realizations or nonunique state space realizations corresponding to a given transfer function (Kailath, 1980). In matrix theory, the transformation (2.2.24) is known as similarity transformation, and A and A are called similar matrices.

2.3 SISO TRANSFER FUNCTION FROM A STATE SPACE REALIZATION Taking the Laplace transformation of equations (2.2.19), y (s ) = c(sI − A)−1 x (0 ) + c(sI − A)−1 bu (s )

(2.3.1)

For the definition of a transfer function, x(0 ) = 0 . Therefore, g (s ) =

y (s ) = c(sI − A)−1 b u (s )

(2.3.2)

cAdj (sI − A)b det(sI − A)

(2.3.3)

Equation 2.3.2 can be expressed as g (s ) = where Adj ( sI − A) = A n−1 + ( s + a1 )A n−2 + … + ( s n−1 + a1s n−2 + … + an−1 )I n

(2.3.4)


12


and det( sI n − A) = s n + a1s n−1 + a2 s n−2 + … + an−1s + an

(2.3.5)

Further, it can be shown that cAdj ( sI − A)b = b1s n−1 + b2 s n−2 + … + bn−1s + bn

(2.3.6)

The results (2.3.5) and (2.3.6) are also true for A, b , and c . Hence, the transfer function is unique for all similar state space realizations.

2.4 SOLUTION OF STATE SPACE EQUATIONS 2.4.1 HOMOGENEOUS EQUATION Consider the solution of Equation 2.2.19 with u (t ) = 0 ; i.e., dx = Ax ; dt

x (0 ) = x 0

(2.4.1)

Taking the Laplace transformation of (2.4.1), x (s ) = (sI − A)−1 x 0

(2.4.2)

−1 ⎞ 1⎛ 1 ⎛ A A2 A⎞ ( sI − A)−1 = ⎜ I − ⎟ = ⎜ I + + 2 + …⎟ s⎝ s⎠ s⎝ s s ⎠

(2.4.3)

Now,

Taking the inverse Laplace transformation of (2.4.3), L−1 ( sI − A)−1 = I + At + A 2

t2 + …; t ≥ 0 2

(2.4.4)

The matrix exponential is defined as follows: e At = L−1[(sI − A)−1 ] Lastly, taking the inverse Laplace transformation of (2.4.2),

(2.4.5)



x (t ) = e At x 0 ;

13

t≥0

(2.4.6)

The matrix e At is also known as the state transition matrix because it takes the initial state x 0 to a state x(t ) in time t. Properties of eAt

1.

d At (e ) = Ae At = e At A dt

(2.4.7)

2. e A (t1+t2 ) = e At1 e At2

(2.4.8)

3. If e At is nonsingular, (e At )−1 = e − At

(2.4.9)

EXAMPLE 2.2 ⎡ −4 A=⎢ ⎣ −1

⎡0 ⎤ 3⎤ ⎥ x(0 ) = ⎢ ⎥ 0⎦ ⎣1⎦

⎡ s ⎢ (s + 1)(s + 3) (sI − A)−1 = ⎢ ⎢ 1 ⎢ − (s + 1)(s + 3) ⎣

⎤ 3 (s + 1)(s + 3) ⎥⎥ ⎥ s+4 (s + 1)(s + 3) ⎥⎦

⎡1.5e −3t − 0.5e − t L−1[(sI − A)−1 ] = e At = ⎢ −3 t −t ⎣0.5e − 0.5e

(2.4.10)

(2.4.11)

−1.5e −3t + 1.55e − t ⎤ ⎥ −0.5e −3t + 1.5e − t ⎦

(2.4.12)

t≥0

(2.4.13)

Then, Equation 2.4.6 yields ⎡ −1.5e −3t + 1.5e − t ⎤ x(t ) = ⎢ ; −3 t −t ⎥ ⎣ −0.5e + 1.5e ⎦

2.4.2 INHOMOGENEOUS EQUATION Taking the Laplace transformation of (2.2.19), x (s ) = (sI − A)−1 x (0 ) + (sI − A)−1 bu (s )

(2.4.14)

Utilizing the definition (2.4.5), the inverse Laplace transformation of (2.4.14) yields


14


t

x (t ) = e x (0 ) + At

∫e

A (t − τ )

bu (τ) d τ ; t ≥ 0

(2.4.15)

0

EXAMPLE 2.3 Let ⎡ −4 A=⎢ ⎣ −1

⎡1⎤ ⎡0 ⎤ 3⎤ ⎥ b = ⎢ ⎥ x(0 ) = ⎢ ⎥ ; and u(t) = 1 for t ≥ 0 0⎦ ⎣0 ⎦ ⎣1⎦ ⎡1.5e −3(t − τ ) − 0.5e − (t − τ ) ⎤ e A (t − τ ) b = ⎢ ⎥ −3 ( t − τ ) − 0.5e − (t − τ ) ⎦ ⎣0.5e

(2.4.16)

(2.4.17)

Then

t

∫ 0

⎡ ⎢1.5 e−3t ⎢ e A (t − τ )bu ( τ )d τ = ⎢ ⎢ −3t ⎢0.5 e ⎢⎣

t

t

∫

e3 τ d τ − 0.5 e−t

t

∫

∫ 0

0

t

e3 τ d τ − 0.5 e−t

0

∫ 0

⎤ eτ d τ ⎥ ⎥ ⎥= ⎥ τ e d τ⎥ ⎥⎦

(2.4.18)

⎡ −0.5 e−3t + 0.5 e−t ⎤ ⎢ ⎥ ⎢− 1 − 0.5 e−3t + 0.5 e−t ⎥ ⎢⎣ 3 3 ⎥⎦ Hence, from (2.4.15), ⎡ −2 e −3t + 2 e − t ⎤ ⎥; x(t ) = ⎢ 1 2 −3t ⎢ − − e + 2e− t ⎥ ⎢⎣ 3 3 ⎥⎦

t≥0

(2.4.19)

2.5 OBSERVABILITY AND CONTROLLABILITY OF A SISO SYSTEM 2.5.1 OBSERVABILITY The state space model {A, b, c} can be developed on the basis of the governing differential Equation 2.1.1 or the transfer function (2.1.3). If the initial conditions



15

on the output and its derivatives are nonzero, how would one get correct values of initial states? To answer this question, consider the output Equation 2.2.19b, y (t ) = cx (t ) Differentiating it and using (2.2.19a), dy = cAx (t ) + cbu (t ) dt

(2.5.1)

Continuing this differentiation process, d2y du = cA 2 x(t ) + cAbu (t ) + cb 2 dt dt

(2.5.2)

d n−1y d n−1u n−1 n −2 n− 3 du = c A x ( t ) + c A b u ( t ) + c A b + … + cb dt n−1 dt n−1 dt Representing (2.5.1) and (2.5.2) in matrix form, y (t ) = Θx (t ) + Τu(t )

(2.5.3)

where ⎡ y(t ) = ⎢ y ⎣

dy dt

.

.

d n −1y ⎤ ⎥ dt n −1 ⎦

T

(2.5.4)

⎡ c ⎤ ⎢ ⎥ ⎢ cA ⎥ ⎢ cA2 ⎥ ⎢ ⎥ Θ=⎢ . ⎥ ⎢ . ⎥ ⎢ ⎥ ⎢ . ⎥ ⎢ n −1 ⎥ ⎣cA ⎦ ⎡ u(t ) = ⎢u ⎣

du dt

.

.

(2.5.5)

d n −1u ⎤ ⎥ dt n −1 ⎦

T

(2.5.6)


16


⎡ 0 ⎢ ⎢ cb ⎢ cAb Τ=⎢ ⎢ . ⎢ . ⎢ n −2 ⎢⎣cA b

0 0 cb . . .

0 0 0 . . .

. . . . . .

. . . . . cb

0⎤ ⎥ 0⎥ 0⎥ ⎥ .⎥ .⎥ ⎥ 0 ⎥⎦

(2.5.7)

From (2.5.3), Θx (0 ) = y (0 ) − Τu(0 )

(2.5.8)

x (0 ) = Θ −1y (0 ) − Θ −1Τu(0 )

(2.5.9)

Therefore,

provided the matrix Θ is nonsingular. Hence, the condition for initial states to be calculated from the initial values of input, output, and their derivatives is that the matrix Θ should be nonsingular. The matrix Θ is known as the observability matrix. The solution (2.4.15) of the state space equation indicates that x(t ) can be calculated for any u(t) if the initial value x(0 ) is known. Observability of State Space Realization Using Method I Using (2.5.5), ⎡ bn − b0 an ⎢ ⎢ − an (b1 − b0 a1 ) Θ=⎢ . ⎢ . ⎢ ⎢ . ⎣

. . . . .

. . . . .

. . . . .

. . . . .

b1 − b0 a1 ⎤ ⎥ . ⎥ ⎥ . ⎥ . ⎥ ⎥ . ⎦

(2.5.10)

The determinant and hence the singularity of the observability matrix will depend on system parameters. Hence, the realization may or may not be observable. Observability of State Space Realization Using Method II Using (2.5.5),



⎡0 ⎢ ⎢0 Θ = ⎢. ⎢ ⎢. ⎢1 ⎣

0 0 . . 0

0 . . . 0

17

. . . . .

. 1 . . .

1⎤ ⎥ 0⎥ .⎥ ⎥ .⎥ 0 ⎦⎥

(2.5.11)

It can be easily seen that det Θ = −1 or + 1

(2.5.12)

Therefore, Θ is nonsingular irrespective of the system parameter values. Hence, this state space realization is always observable. EXAMPLE 2.4 Consider a spring-mass-damper system Figure 2.4, with the following differential equation: mx + αx + βx = f (t )

(2.5.13)

where f(t) is the applied force. Dividing (2.5.13) by m, x +

f (t ) α β x + x = = u (t ) m m m

(2.5.14)

Defining states as x1 = x

and

x2 = x

(2.5.15a,b)

state equations are x = Ax (t ) + bu (t )

(2.5.16)

where ⎡ x (t ) ⎤ x(t ) = ⎢ 1 ⎥ ; ⎣ x2 (t ) ⎦

⎡ 0 A=⎢ ⎣ −β / m

1 ⎤ ⎥; −α / m ⎦

Case I: Position Output If the position of the mass, x, is measured by a sensor,

⎡0 ⎤ b=⎢ ⎥ ⎣1⎦

(2.5.17)


18

Linear Systems: Optimal and Robust Control x (t )

m

f (t )

FIGURE 2.4 A spring-mass-damper system.

y (t ) = cx (t )

(2.5.18)

c = [1

0]

(2.5.19)

⎡1 Θ=⎢ ⎣0

0⎤ ⎥ 1⎦

(2.5.20)

where

Then the observability matrix is

Hence, the state space realization is observable for all values of spring stiffness and damping constant. Case II: Velocity Output If the velocity of the mass, x , is measured by a sensor, y (t ) = cx (t )

(2.5.21)

c = [0

1]

(2.5.22)

1 ⎤ ⎥ −α / m ⎦

(2.5.23)

where

Then the observability matrix is ⎡ 0 Θ=⎢ ⎣ −β / m Therefore,



19

det Θ = β / m

(2.5.24)

Hence, the state space realization is not observable if the spring stiffness is zero, i.e., if there is no spring in the system. This is an interesting result because the displacement can be obtained by integrating velocity: t

x1 (t ) = x1 (0 ) +

∫ x (t)dt

(2.5.25)

2

0

However, the initial condition x1 (0 ) is needed. In the absence of a spring, the loss of observability implies that x1 (0 ) cannot be obtained, and the position of the system cannot be observed.

2.5.2 CONTROLLABILITY The linear differential equation dx = Ax + bu (t ) dt

(2.5.26)

is controllable if and only if it can be transferred from any initial state to any final state in a finite time. Therefore, for the linear and time-invariant system (2.5.26), the issue is to find u(t); t0 ≤ t ≤ t f , which will take the system from any x(t0 ) to any x(t f ) in a finite time t f − t0 . Using the solution of (2.5.26), tf

x (t f ) = e

A ( t f − t0 )

x (t0 ) +

∫e

A (t f − τ )

bu (τ) d τ

(2.5.27)

t0

This is an integral equation because the unknown function u(.) appears under an integral sign. Define a matrix P (t f , t0 ) as follows: tf

P (t f , t0 ) =

∫

e

A (t f − τ )

bb T e

AT ( t f − τ )

dτ

t0

This matrix P (t f , t0 ) is known as the controllability Gramian. The solution of (2.5.27) is found (Friedland, 1985) to be

(2.5.28)


20


u (τ) = b T e

AT ( t f − τ )

P −1 (t f , t0 )[ x (t f ) − e

A ( t f − t0 )

x (t0 )]

(2.5.29)

This result can be verified by substituting (2.5.29) into (2.5.27). Hence, the control input u(.) can be found if and only if the matrix P (t f , t0 ) is nonsingular. Defining a new variable ν = t f − t0 , t f − t0

∫e

P (t f , t0 ) =

Aν

T

bb T e A ν d ν

(2.5.30)

0

Note that P (t f , t0 ) = P (t f − t0 ) ; i.e., P (t f , t0 ) only depends on t f − t0 for a linear and time-invariant system. A t e s t f o r l i n e a r i n d e p e n d e n c e o f a n y f u n c t i o n s { i (τ); 0 ≤ τ ≤ t f − t0 , i = 1, 2,...., n} is that their Gramian matrix G (Appendix A) is nonsingular where t f − t0

G=

∫

T

(τ) (τ) d τ

T

(τ) = ⎡⎣ 1 (τ)

and

.

.

.

n (τ) ⎤⎦

(2.5.31)

0

Equation 2.5.30 can be written as t f − t0

P (t f , t0 ) =

∫e

Aν

b(e Aν b)T d ν

(2.5.32)

0

Comparing (2.5.31) and (2.5.32), nonsingularity of the matrix P implies that elements of the vector e Aν b are linearly independent over (0, t f − t0 ) . Therefore, the pair {A, b} is controllable over (0, t f − t0 ) if and only if the elements of the vector e Aν b are linearly independent. EXAMPLE 2.5 ⎡0 A=⎢ ⎣ −4

1⎤ ⎥ 0⎦

and

⎡0 ⎤ b=⎢ ⎥ ⎣1⎦

(2.5.33)

Find e At : ⎡s sI − A = ⎢ ⎣4

−1⎤ ⎥ s⎦

(2.5.34)



21

⎡ s ⎢ 2 (sI − A) = ⎢ s + 4 ⎢ −4 ⎢ s2 + 4 ⎣ −1

1 ⎤ ⎥ s2 + 4 ⎥ s ⎥ s 2 + 4 ⎥⎦

⎡ cos 2 t e At = L−1[(sI − A)−1 ] = ⎢ ⎣ −2 sin 2 t

0.5 sin 2 t ⎤ ⎥ cos 2 t ⎦

(2.5.35)

(2.5.36)

Find e Aν b : ⎡ sin 2 ν ⎤ sin 2 ν ⎤ ⎡0 ⎤ ⎢ ⎥ = 2 ⎥ ⎢ ⎥ ⎢ 2 ⎥⎥ 1 cos 2 ν ⎥⎦ ⎣ ⎦ ⎢⎣ cos 2 ν ⎥⎦

⎡ cos 2 ν e Aν b = ⎢ ⎢ ⎢⎣ −2 sin 2 ν

(2.5.37)

From the controllability Gramian,

Aν

T AT ν

e bb e

⎡ sin 2 2 ν ⎢ 4 =⎢ ⎢ sin 2 ν cos 2 ν ⎢⎣ 2

sin 2 ν cos 2 ν ⎤ ⎥ 2 ⎥ cos 2 2ν ⎥⎥ ⎦

(2.5.38)

Let t0 − t f = Δt

(2.5.39)

Hence, Δt

P (t f , t0 ) =

∫e

Aν

T

bb T e A ν d ν

0

⎡ Δt − (sin( 4 Δt )) / 4 ⎢ 8 =⎢ − cos( 4 Δt ) + 1 ⎢ ⎢⎣ 16

det( P ( Δt )) =

− cos( 4 Δt ) + 1 ⎤ ⎥ 16 ⎥ Δt + (sin( 4 Δt )) / 4 ⎥ ⎥⎦ 2

( Δt )2 − 0.25 sin 2 (2 Δt ) 16

(2.5.40)

(2.5.41)


22


Theorem The system is completely controllable if and only if the controllability matrix C = [b

Ab

A2 b

.

.

A n −1 b]

(2.5.42)

is nonsingular. Proof Step I: Consider that the matrix P, Equation 2.5.28, is singular. In this case, elements of the vector e Aν b are linearly dependent. As a result, there exists a nonzero vector q such that T

z(t ) = b T e A t q = 0 for 0 ≤ t ≤ t f − t0

(2.5.43)

Differentiating (2.5.43), T

z(t ) = b T AT e A t q = 0 T

z (t ) = bT ( AT )2 e A t q = 0

(2.5.44)

T d n−1 z(t ) = bT ( AT )n−1 e A t q = 0 n−1 dt

Putting (2.5.43) and (2.5.44) in matrix form, ⎡ ⎤ bT ⎢ ⎥ T T ⎢ b A ⎥ T ⎢ ⎥ eA tq = 0 . ⎢ ⎥ . ⎢ ⎥ ⎢b T ( AT ) n −1 ⎥ ⎣ ⎦

(2.5.45)

Using the definition (2.5.42), T

C T eA tq = 0

(2.5.46)



23

Let C T = ⎡⎣r1

r2

.

.

rn ⎤⎦

(2.5.47)

and ⎡ α1 (t ) ⎤ ⎢ ⎥ ⎢ α 2 (t ) ⎥ T eA tq = ⎢ . ⎥ ⎢ ⎥ ⎢ . ⎥ ⎢α (t ) ⎥ ⎣ n ⎦

(2.5.48)

Hence, from (2.5.46) – (2.5.48), α1 (t )r1 + α 2 (t )r2 + … + α n (t )rn = 0

(2.5.49)

Hence, columns of C T are linearly dependent. As rank (C ) = rank (C T ), the matrix C is singular. Step II: Consider that C is singular. We will show that the matrix P, Equation 2.5.28, is singular in this case. e Aν = I + A ν +

A 2 ν2 A n−1νn−1 A n νn +… + + +… 2! (n − 1)! n!

(2.5.50)

Recall the Cayley–Hamilton theorem: A n = −a1A n−1 − a2 A n−2 − … − an I n

(2.5.51)

det( sI − A) = s n + a1s n−1 + a2 s n−2 + … + an

(2.5.52)

where

Using (2.5.50) and (2.5.51), e Aν = If1 ( ν) + Af2 ( ν) + … + A n−1 fn ( ν) where f1 ( ν),.…, fn ( ν) are functions of ν .

(2.5.53)


24


Hence, e Aν b = Cf ( ν)

(2.5.54)

⎡ f1 ( ν) ⎤ ⎢ ⎥ ⎢ f2 ( ν) ⎥ f ( ν) = ⎢ . ⎥ ⎢ ⎥ ⎢ . ⎥ ⎢ f ( ν) ⎥ ⎣ n ⎦

(2.5.55)

where

Therefore, from (2.5.30), t f − t0

P (t f , t0 ) = C [

∫ f (ν)f

T

( ν) d ν]C T

(2.5.56)

0

Note that rank ( AB) ≤ min{rank ( A), rank ( B)} . Therefore, the rank of P (t f , t0 ) is going to be less than n because it has been assumed that the matrix C is singular. Controllability of the State Space Realization Obtained Using Method I Using (2.5.42), ⎡0 ⎢ ⎢0 ⎢. C=⎢ ⎢. ⎢. ⎢ ⎢⎣ 1

0 0 . . 1 − a1

0 . . . − a1 − a 2 + a1

. . . . . .

. 1 . . . .

1⎤ ⎥ 0⎥ .⎥ ⎥ .⎥ .⎥ ⎥ . ⎥⎦

(2.5.57)

It can be shown that det C = −1 or + 1

(2.5.58)

Therefore, the state space realization obtained using Method I is always controllable.



25

Controllability of the State Space Realization Obtained Using Method II Using (2.5.42), ⎡ bn − b0 an ⎢ ⎢ − an (b1 − b0 a1 ) T C =⎢ . ⎢ . ⎢ ⎢ . ⎣

. . . . .

. . . . .

. . . . .

. . . . .

b1 − b0 a1 ⎤ ⎥ . ⎥ ⎥ . ⎥ . ⎥ ⎥ . ⎦

(2.5.59)

The determinant and hence the singularity of the controllability matrix will depend on system parameters. Hence, the realization may or may not be controllable. EXAMPLE 2.6 Consider the tank system shown in Figure 2.5, where u1 and u2 are input mass flow rate to tank 1 and tank 2, respectively. Let p1 and pa be the pressure at the left end of the pipe and the atmospheric pressure, respectively. Then, p1 − pa = ρgh1

(2.5.60)

where ρ is the fluid density. u1 pa

h1

Tank 1

p1

pa

q

h2

FIGURE 2.5 A tank-pipe system.

u2

Tank 2


26


Considering a laminar pipe flow, q=

p1 − pa RL

(2.5.61)

where q and RL are the volumetric flow rate through the pipe and the flow resistance, respectively. From the law of conservation of mass, ρA1

dh1 = u1 − ρq dt

(2.5.62)

ρA2

dh2 = u2 + ρq dt

(2.5.63)

where A1 and A2 are cross-sectional areas of tanks 1 and 2, respectively. Using (2.5.60) and (2.5.61), ⎡ h1 ⎤ ⎡ −α1 ⎢ ⎥=⎢ ⎣ h2 ⎦ ⎣ α 2

⎡0⎤ 0 ⎤ ⎡ h1 ⎤ ⎡β1 ⎤ ⎥ ⎢ ⎥ + ⎢ ⎥ u1 + ⎢ ⎥ u2 0 ⎦ ⎣ h2 ⎦ ⎣ 0 ⎦ ⎣β 2 ⎦

(2.5.64)

where α1 =

ρg , A1RL

α2 =

ρg , A2 RL

β1 =

1 , ρA1

and

β2 =

1 ρA2

(2.5.65)

The controllability matrix with respect to input u1 is ⎡β C=⎢ 1 ⎣0

−α1β1 ⎤ ⎥ α 2 β1 ⎦

(2.5.66)

Hence, the system (2.5.64) is controllable with respect to the input u1. It is obvious that both states h1 and h2 are influenced by the input u1. And, the controllability matrix with respect to the input u2 is ⎡0 C=⎢ ⎣β 2

0⎤ ⎥ 0⎦

(2.5.67)

Hence, the system (2.5.64) is not controllable with respect to the input u2 . It is obvious that the state h1 cannot be influenced by the input u2 .



27

m

f (t )

z1

c

c

Fixed z3

m m

z2 c

FIGURE 2.6 A periodic spring-mass system.

EXAMPLE 2.7: A PERIODIC STRUCTURE For the system in Figure 2.6, differential equations of motion are mz1 + βz1 + βc ( z1 − z3 ) + βc ( z1 − z2 ) = f (t )

(2.5.68)

mz2 + βz2 + βc ( z2 − z1 ) + βc ( z2 − z3 ) = 0

(2.5.69)

mz3 + βz3 + βc ( z3 − z2 ) + βc ( z3 − z1 ) = 0

(2.5.70)

Define x = [ z1

z2

z3

z1

z2

z3 ]T

(2.5.71)

and

u (t ) =

f (t ) m

(2.5.72)

Then state equations are x = Ax (t ) + bu (t ) where

(2.5.73)


28


⎡0 ⎢ ⎢0 ⎢0 A=⎢ ⎢− p ⎢q ⎢ ⎢⎣ q

0 0 0

0 0 0

q −p q

q q −p

p=

1 0 0 0 0 0

β + 2β c m

0⎤ ⎥ 0⎥ 1⎥ ⎥, 0⎥ 0⎥ ⎥ 0 ⎥⎦

0 1 0 0 0 0

and

q=

⎡0 ⎤ ⎢ ⎥ ⎢0 ⎥ ⎢0 ⎥ b=⎢ ⎥ ⎢1⎥ ⎢0 ⎥ ⎢ ⎥ ⎢⎣0 ⎥⎦

2βc m

(2.5.74a,b)

(2.5.75a,b)

The controllability matrix is ⎡0 ⎢ ⎢0 ⎢0 C=⎢ ⎢1 ⎢0 ⎢ ⎢⎣0

1 0 0 0 0 0

0 0 0 −p q q

−p q p 0 0 0

p 2 + 2q 2 ⎤ ⎥ q 2 − 2 pq ⎥ q 2 − 2 pq ⎥ ⎥ 0 ⎥ ⎥ 0 ⎥ ⎥⎦ 0

0 0 0 2 p + 2q 2 q 2 − 2 pq q 2 − 2 pq

(2.5.76)

Therefore, ⎡1 ⎢ ⎢0 ⎢0 det C = det ⎢ ⎢0 ⎢0 ⎢ ⎣⎢0 ⎡1 ⎢ det C = det ⎢0 ⎢0 ⎣

−p q p 0 0 0 −p q q

p 2 + 2q 2 q 2 − 2 pq q 2 − 2 pq 0 0 0

0 0 0 1 0 0

0 0 0 −p

⎡1 p 2 + 2q 2 ⎤ ⎥ ⎢ 2 q − 2 pq ⎥ . det ⎢0 ⎢0 q 2 − 2 pq ⎥⎦ ⎣

−p q q

q q

⎤ 0 ⎥ 0 ⎥ ⎥ 0 ⎥ 2 2 p + 2q ⎥ q 2 − 2 pq ⎥⎥ q 2 − 2 pq ⎦⎥

(2.5.77)

p 2 + 2q 2 ⎤ ⎥ q 2 − 2 pq ⎥ = 0 (2.5.78) q 2 − 2 pq ⎥⎦

In other words, the state space realization is not controllable. Because of symmetry, the influence of input on states z2 and z3 is identical. As a result, the input cannot take the system from any initial states to any arbitrary final states.



29

2.6 SOME IMPORTANT SIMILARITY TRANSFORMATIONS 2.6.1 DIAGONAL FORM Consider the similarity transformation (2.2.24). Find a nonsingular matrix T such that A = Ad = diag(λ1, λ 2 ,…, λ n )

(2.6.1)

T −1 AT = Ad

(2.6.2)

AT = TAd

(2.6.3)

In other words,

or

Define T = ⎡⎣t1

t2

.

.

t n ⎤⎦

(2.6.4)

where t i is the ith column of the matrix T. Hence, from (2.6.3) and (2.6.4), At i = λ i t i

(2.6.5)

Therefore, λ i and t i must be the eigenvalue and the corresponding eigenvector of the matrix A, respectively. The matrix T will be nonsingular if and only if A has n independent eigenvectors. In this context, the following two properties are stated: If A has n distinct eigenvalues, there exists n independent eigenvectors. If an eigenvalue of A is repeated r times, r independent eigenvectors corresponding to this eigenvalue will be found provided: rank ( A − λ i I ) = n − r

(2.6.6)

⎡5 ⎢ A = ⎢ −1 ⎢ −1 ⎣

(2.6.7)

EXAMPLE 2.8 −1 5 −1

−1⎤ ⎥ −1⎥ 5 ⎥⎦


30


Eigenvalues of the matrix A are 3, 6, and 6; i.e., the eigenvalue 6 is repeated twice. ⎡ −1 ⎢ A − 6 I = ⎢ −1 ⎢ −1 ⎣

−1⎤ ⎥ −1⎥ −1⎥⎦

−1 −1 −1

(2.6.8)

Therefore, rank ( A − 6 I ) = 1, the condition (2.6.6) is satisfied, and the matrix A can be digonalized.

2.6.2 CONTROLLABILITY CANONICAL FORM Given a realization { A, b, c} , find the similarity transformation matrix T such that Ac = T −1AT ,

b c = T −1b,

cc = cT

and

(2.6.9a,b,c)

Note that Equation 2.2.29 has been used here. From (2.6.9a), TAc = AT

(2.6.10)

Let T = ⎡⎣t1

t2

.

t n ⎤⎦

.

(2.6.11)

Now, AT = [ At1

At 2

At 3

.

.

At n ]

(2.6.12)

and using (2.2.9), TAc = [− an t n

t1 − an −1t n

t 2 − an −2 t n

From (2.6.10), (2.6.12), and (2.6.13), At1 = − an t n At 2 = t1 − an −1t n At 3 = t 2 − an−2 t n

.

.

t n −1 − a1t n ] (2.6.13)



31

At n = t n −1 − a1t n As b c = [0

0

.

.

0

(2.6.14)

1]T , Equation 2.6.9b yields tn = b

(2.6.15)

Solving equations (2.6.14) from the last equation backward, t n −1 = ( A + a1I )b = Ab + a1b t n −2 = A( A + a1I )b + a2 b = A2 b + a1 Ab + a2 b

(2.6.16)

t 2 = A n−2 b + a1A n−3b + ..... + an−2 b t1 = A n −1b + a1 A n −2 b + ..... + an −1b Representing (2.6.16) in the matrix form, T = CQ

(2.6.17)

where ⎡ an −1 ⎢ ⎢ an −2 Q=⎢ . ⎢ ⎢ a1 ⎢ 1 ⎣

an −2 an −3 . 1 0

. . . 0 0

. . . . .

a1 1 . . .

1⎤ ⎥ 0⎥ .⎥ ⎥ 0⎥ 0 ⎥⎦

(2.6.18)

The matrix Q is always nonsingular. Therefore, the matrix T is nonsingular if and only if the controllability matrix C is nonsingular. In summary, any given realization can be converted to the canonical form {Ac , b c , c c } , provided the controllability matrix is nonsingular.

2.7 SIMULTANEOUS CONTROLLABILITY AND OBSERVABILITY The realization {Ac , b c , c c } is always controllable. On the other hand, the realization {Ao , b o , c o } is always observable. Under what conditions are {Ac , b c , c c } and


32


{Ao , b o , c o } observable and controllable, respectively? To answer this question, consider the following theorems (Kailath, 1980). Theorem 1 A realization is both observable and controllable if and only if there are no common factors between the numerator and the denominator of the transfer function. Theorem 2 Observability and controllability properties are preserved under similarity transformations. Therefore, {Ac , b c , c c } is guaranteed to be observable provided there are no common factors between the numerator and the denominator of the transfer function. Similarly, {Ao , b o , c o } is guaranteed to be controllable provided there are no common factors between the numerator and the denominator of the transfer function. Furthermore, if one can find one realization that is both observable and controllable, all realizations are both observable and controllable. If there are no common factors between numerator and denominator, any realization can be converted to either {Ac , b c , c c } or {Ao , b o , c o }.

2.7.1 OBSERVABILITY METHOD I

OF

STATE SPACE REALIZATION USING

Without any loss of generality, assume that b0 = 0 . Therefore, cc = ⎡⎣bn

bn −1

.

b1 ⎤⎦

.

(2.7.1)

Let ei be the ith row of the identity matrix I n . Then, it can be easily shown (Kailath, 1980) that ei Ac = ei+1 ; e n Ac = ⎡⎣ − an

1≤ i ≤ n −1

− an −1

.

.

(2.7.2) − a1 ⎤⎦

(2.7.3)

Consider b( Ac ) = b1Acn−1 + b2 Acn−2 + … + bn−1Ac + bn I n

(2.7.4)

e1b( Ac ) = b1e1Acn−1 + b2 e1Acn−2 + … + bn−1e1Ac + bn e1I n

(2.7.5)

Then,

Because of (2.7.2),



33

e1 Ac = e2 e1Ac2 = e 2 Ac = e 3

(2.7.6)

e1Acn−1 = e n From (2.7.5) and (2.7.6), e1b( Ac ) = b1e n + b2 e n−1 + … + bn−1e 2 + bn e1 = c c

(2.7.7)

From (2.7.2) and (2.7.7), e2 b ( Ac ) = e1 Ac b ( Ac ) = e1b ( Ac ) Ac = cc Ac

(2.7.8)

e3 b ( Ac ) = cc Ac2 ,…, e n b ( Ac ) = cc Acn −1

(2.7.9)

In a similar manner,

Therefore, the observability matrix is ⎡ c c ⎤ ⎡ e1 ⎤ ⎢ ⎥ ⎢ ⎥ ⎢ cc Ac ⎥ ⎢ e2 ⎥ Θc = ⎢ . ⎥ = ⎢ . ⎥ b ( Ac ) = I n b ( Ac ) = b ( Ac ) ⎢ ⎥ ⎢ ⎥ ⎢ . ⎥ ⎢. ⎥ ⎢c A n −1 ⎥ ⎢e ⎥ ⎣ c c ⎦ ⎣ n⎦

(2.7.10)

Note that b ( Ac ) is a polynomial in Ac . Therefore, it can easily be shown that b ( Ac )p = b (λ)p

(2.7.11)

where λ and p are the eigenvalue and the associated eigenvector of Ac. In other words, the eigenvalue and the associated eigenvector of b ( Ac ) are b(λ) and p, respectively. From (2.7.4), b (λ) = b1λ n −1 + b2 λ n −2 + ........ + bn

(2.7.12)

As the determinant of a matrix equals the product of its eigenvalues, Equation 2.7.10 yields


34


n

det(Θc ) = det(b ( Ac )) =

∏ b (λ ) i

(2.7.13)

i =1

Therefore, the determinant of the observability matrix will be zero if and only if λ i , an eigenvalue of Ac , satisfies b (λ i ) = 0 ; i.e., λ i is also a zero of the transfer function. In other words, the state space realization obtained using Method I is observable if and only if all poles and zeros of the transfer function are distinct. EXAMPLE 2.9 The transfer function (2.2.20) can be written as g (s ) = 1 + h (s )

(2.7.14)

where

h (s ) =

−2 s 2 + 2 s + 2 s + 2 s 2 + 4s + 3 3

(2.7.15)

It can be easily seen that {Ac , b c , c c } and {Ao , b o , c o } , Equation 2.2.21 and Equation 2.2.22, correspond to the transfer function h (s ) . Poles of h (s ) are 1, and −0.5 ± 1.6583 j, whereas zeros are 0.618 and 1.618. In other words, there are no common factors between the numerator and denominator. Therefore, any state space realization is guaranteed to be both observable and controllable. As a result, {Ac , b c , c c } will be observable, and {Ao , b o , c o } will be controllable. EXAMPLE 2.10 From Figure 2.7, y (s ) s +1 = u (s ) s (s + 4)

v (s)

u(s )

s 1 s ( s 4)

r (s )

p s

FIGURE 2.7 A feedback system.

q

(2.7.16)

y (s )



35

r (s ) p , = y (s ) s + q

(2.7.17)

u (s ) = v (s ) − r (s )

(2.7.18)

For the transfer function (2.7.16), Method I yields ⎡ x1 ⎤ ⎡0 ⎢ ⎥=⎢ ⎣ x2 ⎦ ⎣0

1 ⎤ ⎡ x1 ⎤ ⎡0 ⎤ ⎥ ⎢ ⎥ + ⎢ ⎥ u (t ) −4 ⎦ ⎣ x2 ⎦ ⎣ 1 ⎦

(2.7.19)

⎡x ⎤ 1⎤⎦ ⎢ 1 ⎥ ⎣ x2 ⎦

(2.7.20)

y (t ) = ⎡⎣1 And, for the transfer function (2.7.17),

r + qr (t ) = py (t )

(2.7.21)

Combining (2.7.19) and (2.7.21) and using (2.7.18) and (2.7.20), the state space realization is obtained: ⎡ x1 ⎤ ⎡ 0 ⎢ ⎥ ⎢ ⎢ x2 ⎥ = ⎢ 0 ⎢ r ⎥ ⎢ p ⎣ ⎦ ⎣

1 −4 p

y (t ) = [1

0 ⎤ ⎡ x1 ⎤ ⎡0 ⎤ ⎥⎢ ⎥ ⎢ ⎥ −1 ⎥ ⎢ x2 ⎥ + ⎢ 1 ⎥ v (t ) − a ⎥⎦ ⎢⎣ r ⎥⎦ ⎢⎣0 ⎥⎦

1

(2.7.22)

⎡ x1 ⎤ ⎢ ⎥ 0 ] ⎢ x2 ⎥ ⎢r ⎥ ⎣ ⎦

(2.7.23)

⎤ −4 ⎥ 16 − p ⎥ p − 4 p − pq ⎥⎦

(2.7.24)

Controllability: ⎡0 ⎢ C = ⎢1 ⎢0 ⎣

1 −4 p

det C = p (1 − q ) The realization is not controllable when p = 0 or q = 1 .

(2.7.25)


36


Observability: ⎡ 1 ⎢ O=⎢ 0 ⎢− p ⎣

0 ⎤ ⎥ −1 ⎥ 3 + q ⎥⎦

1 −3 12 − p

det O = 3(1 − q )

(2.7.26)

(2.7.27)

The realization is not observable when q = 1 . Note that y (s ) (s + 1)(s + q ) = v (s ) s (s + 3)(s + q ) + p (s + 1)

(2.7.28)

When p = 0 or q = 1, (s + q ) or (s + 1) is a common factor between the numerator and denominator of the transfer function (2.7.28), respectively. Therefore, a state space realization cannot be both observable and controllable when p = 0 or q = 1. Here, the realization is not controllable when p = 0 , and is neither observable nor controllable when q = 1 .

2.8 MULTIINPUT/MULTIOUTPUT (MIMO) SYSTEMS The transfer function matrix of a MIMO system is defined as y (s ) = G (s )u(s )

(2.8.1)

where y(s ) and u(s ) are px1 and mx1 vectors, respectively. ⎡ y1 (s ) ⎤ ⎢ ⎥ ⎢ y2 (s ) ⎥ y(s ) = ⎢ . ⎥ ⎢ ⎥ ⎢ . ⎥ ⎢ y (s ) ⎥ ⎣ p ⎦

and

⎡ u1 (s ) ⎤ ⎢ ⎥ ⎢ u2 (s ) ⎥ u(s ) = ⎢ . ⎥ ⎢ ⎥ ⎢ . ⎥ ⎢u (s ) ⎥ ⎣ m ⎦

(2.8.2a,b)

Accordingly, the matrix G (s ) is of order pxm. As an example, for a 2-input/2-output system, ⎡ 1 ⎡ y1 (s ) ⎤ ⎢ s + 1 ⎢ ⎥=⎢ ⎣ y2 (s ) ⎦ ⎢ s + 5 ⎢ (s + 1)2 ⎣

1 ⎤ s + 2 ⎥⎥ ⎡ u1 (s ) ⎤ 1 ⎥ ⎢⎣u2 (s ) ⎥⎦ s + 3 ⎥⎦

(2.8.3)



37

Taking the least common multiple of denominators of elements of the matrix G (s ) , G (s ) =

1 N (s ) (s + 1) (s + 2 )(s + 3)

(2.8.4)

2

where ⎡ (s + 1)(s + 2 )(s + 3) N (s ) = ⎢ ⎣ (s + 2 )(s + 3)(s + 5)

(s + 1)2 (s + 3) ⎤ ⎥ (s + 1)2 (s + 2 ) ⎦

(2.8.5)

The elements of the matrix N (s ) are polynomials in s. Hence, N (s ) will be called a polynomial matrix. It can also be expressed as follows: N (s ) = s 3 N1 + s 2 N 2 + sN 3 + N 4

(2.8.6)

where ⎡1 N1 = ⎢ ⎣1

1⎤ ⎥; 1⎦

⎡6 N2 = ⎢ ⎣10

5⎤ ⎥; 4⎦

⎡11 N3 = ⎢ ⎣31

7⎤ ⎥; 5⎦

⎡6 N4 = ⎢ ⎣30

3⎤ ⎥ 2⎦

(2.8.7)

In general, a pxm transfer function matrix G(s) is represented as G (s ) =

N (s ) d (s )

(2.8.8)

where d ( s ) = s r + d1s r−1 + … + dr

(2.8.9)

N ( s ) = s r−1N1 + s r−2 N 2 + … + sN r−1 + N r

(2.8.10)

and

N1 , N 2 ,…, N r are pxm matrices. Definitions 1. A rational transfer function matrix G(s) is said to be proper if lim G (s ) < ∞

s →∞

and strictly proper if

(2.8.11)


38


lim G (s ) = 0

(2.8.12)

s →∞

2. A vector of polynomials is called a polynomial vector. The degree of a polynomial vector equals the highest degree of all the entries of the vector. For example, the polynomial vector ⎡ s4 + 1 ⎤ ⎢ 2 ⎥ ⎢s + s + 1⎥ ⎢ s+7 ⎥ ⎣ ⎦

(2.8.13)

has degree = 4. Lemma If G(s) is a strictly proper (proper) transfer function matrix and G (s ) = N (s ) D −1 (s )

(2.8.14)

then every column (Kailath, 1980) of N(s) has degree strictly less than (less than or equal to) that of the corresponding column of D(s). Definition: Column-Reduced Matrix Let ki = the degree of ith column of m × m matrix D(s)

(2.8.15)

Then, it can be easily seen that m

deg det D (s ) ≤

∑k

i

(2.8.16)

i=1

Inequality holds when there are cancellations of terms in the expansion of det D(s). A matrix D(s) for which the equality sign holds is called a column reduced matrix (Kailath, 1980). EXAMPLE 2.11 ⎡ s4 + s D (s ) = ⎢ 2 ⎣s + s + 1

s2 + 2⎤ ⎥ 1 ⎦

(2.8.17)



39

Here, k1 = 4 and k2 = 2

(2.8.18)

det D (s ) = −s 3 − 3s 2 − s − 2

(2.8.19)

and

In other words, 2

deg det D (s )
0

(2.15.3)

Differentiating (2.15.3) with respect to time, V (t ) = x T (t ) Px (t ) + x T (t ) Px (t )

(2.15.4)

Substituting (2.15.1) into (2.15.4), V (t ) = x T (t )[ AT P + PA]x (t )

(2.15.5)

For a stable system, V (t ) is a negative definite function. Therefore, AT P + PA = −Q

(2.15.6)

where Q = Q T > 0 . Equation 2.15.6 is known as the Lyapunov equation, often written as AT P + PA + Q = 0

(2.15.7)

An Important Property of the Lyapunov Equation When Q = Q T > 0 , and all the eigenvalues of A have negative (nonzero) real parts, the Lyapunov equation 2.15.7 has a unique solution for P, satisfying P = P T > 0 as follows (Kailath, 1980): ∞

P=

∫e 0

AT t

Qe At dt

(2.15.8)



73

EXERCISE PROBLEMS P2.1

Consider the following transfer function:

G (s ) =

P2.2

Develop state space realizations using Methods I and II. Write state space equations and draw the block diagram for each method. Using the properties of the determinant, find the characteristic polynomial of the following matrix: ⎡ 0 ⎢ ⎢ 0 ⎢ . ⎢ ⎢ 0 ⎢−a ⎣ n

P2.3

P2.4

4s 3 + 25s 2 + 45s + 34 s 3 + 6 s 2 + 10 s + 8

1 0 . 0 − an −1

0 1 . 0 .

. . . . .

. . . . .

0 ⎤ ⎥ 0 ⎥ . ⎥ ⎥ 1 ⎥ − a1 ⎥⎦

a. If A = T −1 AT , show that e At = Te At T −1 . b. If there are n independent eigenvectors for the n × n matrix A, develop an algorithm to compute e At using the result in part (a). Matlab Exercise Consider the system shown in the Figure P.2.4 where +

G(s)

P(s)

H(s)

FIGURE P.2.4 A Feedback System.

G (s ) =

s+2 2 s + 6s + 8

P (s ) =

H (s ) =

s 3 + 2 s 2 + 9s + 10 s 3 + s 2 + 10 s + 20

s +1 s+5

a. Develop state space models for both open-loop and closed-loop systems.


74

P2.5


b. Examine controllability and observability of these state space models. c. Determine the stability of open- and closed-loop systems. Consider a single input/single output system with the following transfer function: s+α (s + 1)(s + 2)(s + 3)

P2.6

a. Construct a state space realization {A, b, c} that is controllable for all values of α . You should also show the corresponding simulation diagram. b. Find the values of α for which the state space realization developed in part (a) is not observable. Consider a linear system described by the following state space equation: dx1 = x2 dt dx2 = − x1 + u dt y (t ) = x1 (t ) + x2 (t ) dy (0 ) = 1, and u(0) = 0. dt a. Determine x1 (0 ) and x2 (0 ). b. Determine the response x1 (t ) and x2 (t ) via the matrix exponential when u(t) =1 for t > 0. Consider a single input/single output system with the following transfer function: Given y(0) = 0,

P2.7

s+2 (s + 1)(s + 2 )(s + 3) a. Construct a state space realization {A, b, c} that is controllable. You should also show the corresponding simulation diagram. b. Answer the following questions: i. Will it be possible to construct a realization that is both controllable and observable? ii. Will it be possible to construct a realization that is neither controllable nor observable?



P2.8

75

Consider the following system: dx = Ax (t ) + bu (t ) dt where ⎡0 A=⎢ ⎣0

1⎤ ⎥, 0⎦

⎡0 ⎤ ⎡0 ⎤ b = ⎢ ⎥ , and x(0 ) = ⎢ ⎥ ⎣1⎦ ⎣0 ⎦

Find u(t), 0 ≤ t ≤ 1 , such that x(1) = [1 0 ]T . P2.9 Consider the state space realization using Method II. Show that this realization is controllable, provided there is no common factor between numerator and denominator of the transfer function. P2.10 Consider the following state space realization: dx ⎡ −2 =⎢ dt ⎣ 1

⎡1⎤ 1⎤ ⎥ x (t ) + ⎢ ⎥ u (t ) −2 ⎦ ⎣0 ⎦

y (t ) = ⎡⎣1

2 ⎤⎦ x (t )

i. Find the transfer function of the system. ii. Determine the observability and controllability of the state space realization. iii. Let y(0) = 1,

dy (0 ) = 0, and u(0) =0. Find x(0). dt

iv. Find e At . v. Solve the state equations when x(0) = 0 and u(t) = 1 for t ≥ 0 . P2.11 Consider the following transfer function matrix: ⎡ s +1 H(s) = ⎢ 2 ⎣s + s +1 a. Develop the state controllability and b. Develop the state controllability and

⎤ s+2 ⎥ 2 s + 2s + 1 ⎦ 3

space realization using Method I. Determine its observability. space realization using Method II. Determine its observability.


76


P2.12 Consider the problem of the control of a tire-tread extrusion line: ⎡ 0.071 ⎢ s + 0.19 G (s ) = ⎢ ⎢ 0.24 ⎢ 2 ⎣ (s + 0.23)

0.007 (s + 0.23)2 0.027 (s + 0.3)2

⎤ ⎥ ⎥ ⎥ ⎥ ⎦

a. Find the controller-form state space realization using a suitable MFD. b. Determine controllability and observability of your state space realization and discuss your results. c. Find poles and zeros of the system. P2.13 Consider a system with the following transfer function matrix: ⎡ s ⎢ − (s + 1) G(s) = ⎢ ⎢ (2 s + 1) ⎢ s (s + 1) ⎣

1 ⎤ (s + 1) ⎥⎥ 1 ⎥ s + 1 ⎥⎦

i. Find an irreducible right MFD. ii. What is the order of a minimal realization? iii. Find the poles and transmission zeros. P2.14 Consider the linearized dynamics of a spark ignition engine (Abate et al., 1994): ⎡ y1 (s ) ⎤ α1α 4 1 ⎡ ⎢ ⎥= ⎢ ⎣ y2 (s ) ⎦ d (s ) ⎣ α1 ( Js + α 7 − α 5 )

α 6 (s + α 2 ) ⎤ ⎡ u1 (s ) ⎤ ⎥⎢ ⎥ −α 3 α 6 ⎦ ⎣u2 (s ) ⎦

where y1 (t ) = engine speed, y2 (t ) = relative air pressure of manifold, u1 (t ) = duty cycle of the throttle valve, and u2 (t ) = spark advance position. The polynomial d(s) is defined as follows: d (s ) = Js 2 + ( J α 2 + α 7 − α 5 )s + (α 3α 4 + α 2 α 7 − α 2 α 5 ) Parameters J (mass-moment of inertia) and α1 − α 7 are provided in Table 2.P.1 for three different operating conditions I, II, and III.



77

TABLE 2.P1 Parameters of a Spark Engine

I II III

J

α1

α2

α3

α4

α5

α6

α7

1 1 10

2.1608 3.4329 2.1608

0.1027 0.1627 0.1027

0.0357 0.1139 0.0357

0.5607 0.2539 0.5607

2.0183 1.7993 1.7993

4.4962 2.0247 4.4962

2.0283 1.8201 1.8201

a. Develop a minimum-order state space realization. b. Find poles and zeros of the transfer function matrix for each operating condition.


9217_C003.fm Page 79 Thursday, December 21, 2006 7:44 AM

3

State Feedback Control and Optimization

First, the design of a full state feedback control system is presented for a single input/single output (SISO) system along with its impact on poles and zeros of the closed-loop system. Next, the full state feedback control system is presented for a multiinput/multioutput (MIMO) system. The necessary conditions for the optimal control are then derived and used to develop the linear quadratic (LQ) control theory and the minimum time control.

3.1 STATE VARIABLE FEEDBACK FOR A SINGLE INPUT SYSTEM Consider the state space realization {A, b, c} ; i.e., dx = Ax (t ) + bu (t ) dt

(3.1.1)

y (t ) = cx (t )

(3.1.2)

The open-loop transfer function is given as cAdj (sI − A)b y (s ) = c(sI − A)−1 b = u (s ) det(sI − A)

(3.1.3)

The state variable feedback system is shown in Figure 3.1, in which u (t ) = v (t ) − kx (t )

(3.1.4)

where v (t ) is the external input and k is the state feedback vector. Using (3.1.1) and (3.1.4), dx = ( A − bk ) x (t ) + bv (t ) dt

(3.1.5)

79


80


v (t )

u(t )

x (t ) x

Ax (t )

bu (t )

y (t ) c

k

FIGURE 3.1 A state feedback control system.

Equation 3.1.5 represents the state dynamics of the closed-loop system. The transfer function of the closed-loop system is y (s ) cAdj (sI − A + bk )b = c(sI − A + bk )−1 b = v (s ) det(sI − A + bk)

(3.1.6)

It is desirable to study the effects of state feedback on poles and zeros of the closed-loop transfer function.

3.1.1 EFFECTS OF STATE FEEDBACK ON POLES OF THE CLOSED-LOOP TRANSFER FUNCTION: COMPUTATION OF STATE FEEDBACK GAIN VECTOR It will be shown here that poles of the closed-loop transfer function can be located anywhere in the s-plane by the full state feedback, provided the state space realization is controllable. In other words, it is always possible to find the state feedback vector k corresponding to any desired locations of closed-loop poles, provided the state space realization is controllable. Let y( s ) b1s n−1 + … + bn−1s + bn = n u ( s ) s + a1s n−1 + a2 s n−2 + … + an−1s + an

(3.1.7)

Hence, the characteristic polynomial of the open-loop system is a( s ) = det( sI − A) = s n + a1s n−1 + a2 s n−2 + … + an−1s + an

(3.1.8)

Let the desired characteristic polynomial of the closed-loop system be α( s ) = det( sI − A + bk ) = s n + α1s n−1 + α 2 s n−2 + … + α n−1s + α n where α1 , α 2 , …, α n are arbitrary real numbers.

(3.1.9)



81

Now, α( s ) = det( sI − A + bk ) = det{( sI − A)( I + ( sI − A)−1 bk )} = a( s ).det( I + ( sI − A)−1 bk )

(3.1.10)

Using the identity det( I n − PQ ) = det( I m − QP ) , where P and Q are n × m and m × n matrices, respectively, Equation 3.1.10 can be written (Kailath, 1980) as α(s ) = a (s )[1 + k (sI − A)−1 b]

(3.1.11)

α(s ) − a (s ) = kadj (sI − A)b

(3.1.12)

or

Using Equation 2.3.4 and equating coefficients of s n−1 , s n−2 , …, s 0 on both sides, we obtain α1 − a1 = kb α 2 − a2 = kAb + a1kb α 3 − a3 = kA 2 b + a1kAb + a2 kb

(3.1.13)

α n − an = kA n−1b + a1kA n−2 b + a2 kA n−3b + … + an−1kb Representing the system of equations in matrix form, α − a = kCUt

(3.1.14)

where C is the controllability matrix, α = [α1

α2

α3

.

.

αn ]

(3.1.15)

a = ⎡⎣ a1

a2

a3

.

.

an ⎤⎦

(3.1.16)

and U t is the upper triangular Toeplitz matrix defined as


82


⎡1 ⎢ 0 Ut = ⎢ ⎢. ⎢ ⎢⎣0

a1 1 . 0

.

a2 a1 . .

a2 . .

. . . 0

an −1 ⎤ ⎥ an −2 ⎥ . ⎥ ⎥ 1 ⎥⎦

(3.1.17)

It can be seen that U t is always nonsingular. Hence, for any choice of α , k can be found, provided C is nonsingular (or the system is controllable). The result is k = (α − a )Ut−1C −1

(3.1.18)

This is also known as the Bass–Gura formula for computing k . The fact that the eigenvalues can be arbitrarily relocated by state feedback is known as modal controllability. This can be understood by realizing that states of a system possess all the information about system dynamics. Hence, feeding back all the states is equivalent to using all the information in deciding the control input. Another important point to note is that the order of the closed-loop system is the same as that of the open-loop system, which is not always true for the output feedback via a compensator. EXAMPLE 3.1: BALANCING

A

POINTER

Consider the balancing of a pointer (Kailath, 1980) on your fingertip (Figure 3.2). If you try to do it yourself, you will find that the pointer will fall down when the fingertip does not move. Furthermore, the fingertip must move to balance the pointer. In fact, the acceleration of the fingertip serves as the control input. To develop a mathematical model, the following assumptions are made (Kailath, 1980): 1. The mass of the pointer is concentrated at the top end. 2. The angle φ is small. 3. The force F from the fingertip is applied only along the direction of the pointer. y

m

mg L

0

FIGURE 3.2 A pointer on the fingertip.

xc

x



83

Applying Newton’s second law of motion, mxc = F sin φ ≈ F φ

(3.1.19)

where xc is the x-coordinate of the center of mass. For a small φ, the acceleration of the center of mass along the y-direction can be neglected. Therefore, mg = F cos φ ≈ F

(3.1.20)

xc = gφ

(3.1.21)

xc (t ) = ξ(t ) + L φ(t )

(3.1.22)

From (3.1.19) and (3.1.20),

Now,

where ξ(t ) is the x-coordinate of the fingertip. Substituting (3.1.22) into (3.1.21), = g φ − u (t ) φ L

(3.1.23)

where

u (t ) =

ξ L

(3.1.24)

The variable u (t ) , which is proportional to the fingertip acceleration ξ , will be treated as the control input. The state variable equations are described as ⎡0 ⎡ q1 ⎤ ⎢ ⎢ ⎥= ⎢g ⎣q2 ⎦ ⎢ ⎣L

1⎤ ⎥ ⎡ q1 ⎤ + ⎡ 0 ⎤ u ⎢ ⎥ ⎢ ⎥ 0 ⎥ ⎣q2 ⎦ ⎣ −1⎦ ⎥⎦

(3.1.25)

where q1 = φ Furthermore,

and

q2 = φ

(3.1.26)


84


⎡0 A= ⎢g ⎢ ⎢⎣ L

1⎤ ⎥ 0⎥ ⎥⎦

⎡0⎤ b=⎢ ⎥ ⎣ −1⎦

and

(3.1.27)

Open-Loop Stability det(sI − A) = s 2 −

g L

(3.1.28)

g g and − . Hence, the open-loop system is unstable L L which is exhibited by the fact that the pointer falls down if the fingertip does not have any acceleration, i.e., u(t) = 0. Furthermore, if L is smaller, the magnitude of Open-loop eigenvalues are

g is larger. This matches with the fact that it is harder L to balance a small pointer. Try it. the unstable eigenvalue

Controllability The controllability matrix is

C = ⎡⎣b

⎡0 Ab ⎤⎦ = ⎢ ⎣ −1

−1⎤ ⎥ 0⎦

(3.1.29)

The matrix C is nonsingular. Hence, the state space realization is controllable. State Feedback Control

u (t ) = − k1q1 − k2 q2 = − ⎡⎣ k1

⎡0 A − bk = ⎢ g ⎢ ⎢⎣ L

1⎤ ⎥ − ⎡ 0 ⎤ ⎡k ⎢ ⎥⎣ 1 0 ⎥ ⎣ −1⎦ ⎥⎦

⎡ s ⎢ det( sI − A + bk ) = det ⎢ ⎛ g ⎞ − ⎜ + k1 ⎟ ⎠ ⎣⎢ ⎝ L

⎡q ⎤ k2 ⎤⎦ ⎢ 1 ⎥ ⎣q2 ⎦

⎡ 0 k2 ⎤⎦ = ⎢ g ⎢ +k 1 ⎢⎣ L

(3.1.30)

1⎤ ⎥ k2 ⎥ ⎥⎦

−1 ⎤ ⎛g ⎞ ⎥ 2 = s − k2 s − ⎜ + k1 ⎟ s − k2 ⎥ ⎝L ⎠ ⎥⎦

(3.1.31)

(3.1.32)



85

Let the desired closed-loop poles be –1 and –2. In this case, the desired closedloop characteristic equation will be det(sI − A + bk ) = (s + 1)(s + 2 ) = s 2 + 3s + 2

(3.1.33)

Matching the coefficients of polynomials, − k2 = 3 ⇒ k2 = −3

(3.1.34)

⎛g ⎞ g − ⎜ + k1 ⎟ = 2 ⇒ k1 = − − 2 ⎝L ⎠ L

(3.1.35)

Comments 1. To determine state feedback vector, the Bass–Gura formula (3.1.18) can also be used. But for a low-order problem, it is more straightforward to directly match the coefficients of polynomials. The Bass–Gura formula is extremely useful to solve higher order problems via a computer software such as MATLAB®. 2. The control input is u (t ) = − k1φ − k2 φ . Its implementation requires measurements of φ and φ , and real-time computation of − k1φ − k2 φ . When a human being tries to balance a pointer, one can say that the eyes are estimating the values of φ and φ , and the brain is deciding on a suitable control input and instructing the motor to move the fingertip with proper acceleration.

3.1.2 EFFECTS OF STATE FEEDBACK ON ZEROS OF THE CLOSED-LOOP TRANSFER FUNCTION If the state space realization is controllable, it can be converted to the controller canonical form {Ac , b c , c c } described in Chapter 2, Section 2.2. Under the state feedback, the system matrix of the closed-loop system is given as ⎡ 0 ⎢ ⎢ 0 Ac − b c k = ⎢ ⎢ ⎢ 0 ⎢− a − k 1 ⎣ n

1 0

0 1

0

0

− an −1 − k2

.

. 0 . .

. . . .

⎤ ⎥ ⎥ ⎥ ⎥ 1 ⎥ − a1 − kn ⎥⎦ 0 0

(3.1.36)

Hence, the realization {Ac − b c k, b c , c c } remains in the controller canonical form. Therefore, the transfer function of the closed-loop system is given as


86


c c ( sI − Ac + b c k )−1 b c =

b1s n−1 + … + b n−1 s + bn s n + (a1 + kn )s n−1 + … + (an + k1 )

(3.1.37)

This expression clearly indicates that the numerator polynomial of the transfer function remains unchanged. In other words, the state feedback has no influence on the zeros of the transfer function (Kailath, 1980). EXAMPLE 3.2: BICYCLE DYNAMICS Consider a simple model of a bicycle (Lowell and McKell, 1982; Astrom et al., 2005), in which the rider, wheels, front-fork assembly, and the rear frame are treated as a single rigid body as shown by the plane in Figure 3.3. The total mass is m and the location of the center of gravity (cg) is shown in the figure. Assume that the forward velocity v of the bicycle is a constant. The small angle θ is the perturbation from the bicycle’s upright position. As the weight mg will further try to increase this angle, the rider turns the handlebar by a small angle α, so that the bicycle begins to travel in a circle with the instant radius r and the instant center of rotation O. This circular travel is represented by the angle φ about the vertical axis passing through the rear-wheel contact point. Therefore, the acceleration of the cg in the direction normal to the plane is + bφ + hθ

v2 r

(3.1.38)

where v 2 / r is the centripetal acceleration, and b is the distance of cg from the vertical axis passing through the rear-wheel contact point. Applying Newton’s second law, 2⎞ ⎛ + bφ + v ⎟ = mgθ m ⎜ hθ r ⎠ ⎝

(3.1.39)

where mgθ is the component of the weight along the direction normal to the frame for small θ. v

cg

mg h Front Wheel Contact Point

a

Rear Wheel Contact Point

r O

FIGURE 3.3 Fundamentals of bicycle dynamics.



87

From the definition of the instantaneous center for rotation: r φ = v

and

rα = a

(3.1.40a,b)

Substituting (3.1.40) into (3.1.39), the input/output equation is obtained: 2 − g θ = − bv α − v α θ h ha ah

(3.1.41)

where angles θ(t ) and α(t ) are the output and the input, respectively. Taking the Laplace transform of (3.1.41) with zero initial conditions: θ(s ) b s + b2 = 2 1 α(s ) s + a1s + a2

(3.1.42)

where

b1 = −

bv ; ha

b2 = −

v2 ; ha

a1 = 0 ;

a2 = −

g h

(3.1.43)

Now, a state space model can be constructed via either Method I or Method II in Chapter 2, Section 2.2. There are two poles of the system:

−

g h

and

−

g h

(3.1.44)

And there is one zero:

−

v b

(3.1.45)

The open-loop system is unstable, and has a zero in the left half plane. The system can be stabilized by a simple feedback law: α = kpθ Substituting (3.1.46) into (3.1.39), the closed-loop system is

(3.1.46)


88


⎛ 2 ⎞ + bv k pθ + ⎜ v k p − g ⎟ θ = 0 θ ha h⎠ ⎝ ah

(3.1.47)

For stability,

kp >

ag v2

(3.1.48)

Note that the control law (3.1.46) has to be implemented by the rider using his or her eye, brain, or hand. Equation 3.1.48 indicates that the lower bound of k p is smaller at a higher velocity. In fact, the damping is also higher at a higher velocity, Equation 3.1.47. Therefore, the rider finds it easier to stabilize the bicycle at a higher velocity. State Feedback Control Using Method I (Chapter 2, Section 2.2): x = Ax (t ) + bu (t )

(3.1.49)

y (t ) = cx (t )

(3.1.50)

where ⎡ 0 A=⎢ ⎣g / h

1⎤ ⎥; 0⎦

⎡0 ⎤ b=⎢ ⎥ ⎣1⎦

and

c = ⎡⎣b1

b2 ⎤⎦

(3.1.51)

and ⎡ s det(sI − A + bk ) = det ⎢ ⎣ − g / h + k1

−1 ⎤ 2 ⎥ = s + k2 s + ( k1 − g / h ) s + k2 ⎦

(3.1.52)

Let the desired characteristic polynomial be det(sI − A + bk ) = s 2 + α1s + α 2

(3.1.53)

Matching coefficients of polynomials in (3.1.52) and (3.1.53), k1 = α 2 + g / h

(3.1.54)



89

and k2 = α1

(3.1.55)

And, the closed-loop transfer function is c(sI − A + bk )−1 b =

b1s + b2 s + α1s + α 2 2

(3.1.56)

Note that the zero of the transfer function remains unchanged, whereas the poles have been located arbitrarily via state feedback control: u (t ) = ur (t ) − kx (t )

(3.1.57)

where ur (t ) is the reference input, which will be zero in this case. Compared to the state feedback law (3.1.57), the proportional controller (3.1.46) is simpler and hence easier to be implemented by a rider. This may explain why a person can ride a bicycle for hours, whereas he or she can balance a pointer on a finger only for a few minutes as full state feedback (Equation 3.1.30) is necessary for stability.

3.1.3 STATE FEEDBACK CONTROL CONSTANT OUTPUT

FOR A

NONZERO

AND

Let the external input v(t) be a constant v0 . From Equation 3.1.5, the steady state value of the state vector x s is given by 0 = ( A − bk ) x s + bv0

(3.1.58)

x s = −( A − bk )−1 bv0

(3.1.59)

or

Hence, the steady state output is given as ys = cx s = −c( A − bk )−1 bv0

(3.1.60)

If the desired value of the output is y d , the corresponding command input v0 is obtained by setting ys = y d and solving (3.1.60). The result is v0 = −

yd provided c( A − bk )−1 b ≠ 0 c( A − bk )−1 b

(3.1.61)


90


Now, let us examine the conditions under which c( A − bk )−1 b ≠ 0 . Substituting s = 0 in Equation 3.1.6, it is clear that c( A − bk )−1 b ≠ 0 if there is no zero of the closed-loop transfer function at s = 0. As locations of zeros remain unchanged under state feedback, the command input for the nonzero set point can be found, provided the open-loop transfer function does not have any zero located at s = 0 (Kailath, 1980). EXAMPLE 3.3: SPRING-MASS SYSTEM For the spring-mass-damper system (Chapter 2, Figure 2.4), state equations are defined by (2.5.16) and (2.5.17). Case I: Position Output If the position of the mass, x, is measured by a sensor, y (t ) = cx (t )

(3.1.62)

c = [1

(3.1.63)

where 0]

Let the state feedback control law be u (t ) = v0 − kx (t )

(3.1.64)

It can be shown that −1

⎛ β⎞ c( A − bk )−1 b = − ⎜ k1 + ⎟ ⎝ m⎠

(3.1.65)

From Equation 3.1.61, to achieve the desired set point y d for the output, v0 = −

⎛ β⎞ yd = yd ⎜ k1 + ⎟ −1 ⎝ c( A − bk ) b m⎠

(3.1.66)

If f0 is the force corresponding to v0 , v0 = From (3.1.66) and (3.1.67),

f0 m

(3.1.67)



91

f0 = y d (mk1 + β)

(3.1.68)

f (t ) = f0 − mk1 x − mk2 x

(3.1.69)

From (3.1.64),

Substituting (3.1.69) into the differential Equation 2.5.14, mx + (α + mk2 ) x + (β + mk1 ) x = f0

(3.1.70)

Therefore, position and velocity feedback coefficients k1 and k2 add to the stiffness and damping coefficient, respectively. Furthermore, conditions (3.1.66) and (3.1.68) represent the static equilibrium condition. Case II: Velocity Output If the velocity of the mass, x , is measured by a sensor, y (t ) = cx (t )

(3.1.71)

c = [0

(3.1.72)

where 1]

In this case, c( A − bk )−1 b = 0

(3.1.73)

And a nonzero set point y d cannot be achieved. Mathematically, this is due to the presence of a zero at s = 0 as the transfer function is y (s ) ms = u (s ) ms 2 + αs + β

(3.1.74)

Physically, nonzero constant velocity is not possible when the applied force is a constant.

3.1.4 STATE FEEDBACK CONTROL UNDER CONSTANT INPUT DISTURBANCES: INTEGRAL ACTION Consider the following system:


92


dx = Ax + bu (t ) + w dt

(3.1.75)

y (t ) = cx (t )

(3.1.76)

where w is a constant disturbance vector of unknown magnitudes. The objective is to develop a state feedback control algorithm such that lim y(t ) = 0. The reader can t →∞

verify that the algorithm (3.1.4) will not be able to achieve this goal. The required algorithm is developed by defining a new variable q(t) as follows: dq = y (t ) = cx (t ); dt

q(0 ) = 0

(3.1.77)

Defining a new state vector, ⎡ x (t ) ⎤ p(t ) = ⎢ ⎥ ⎣ q (t ) ⎦

(3.1.78)

⎡w⎤ dp = Aa p(t ) + b a u (t ) + ⎢ ⎥ dt ⎣0 ⎦

(3.1.79)

⎡A Aa = ⎢ ⎣c

(3.1.80)

Using (3.1.75) to (3.1.77),

where 0⎤ ⎥; 0⎦

⎡b ⎤ ba = ⎢ ⎥ ⎣0 ⎦

Lemma The augmented system {Aa , b a } is controllable provided that the original system {A, b} is controllable and ⎡A rank ⎢ ⎣c

b⎤ ⎥ = n +1 0⎦

Proof The controllability matrix (Gopal, 1984) for the system {Aa , b a } is

(3.1.81)



⎡b Ca = ⎢ ⎣0

93

A2 b cAb

Ab cb ⎡A = ⎢ ⎣c

. .

b ⎤ ⎡0 ⎥⎢ 0 ⎦ ⎣1

. .

Anb ⎤ ⎥ cA n −1b ⎦

C⎤ ⎥ 0⎦

(3.1.82)

(3.1.83)

The matrix ⎡0 ⎢ ⎣1

C⎤ ⎥ 0⎦

(3.1.84)

is nonsingular if the original system {A, b} is controllable; i.e., the matrix C is nonsingular. Therefore, Ca is nonsingular if the condition (3.1.81) is satisfied. Now, the state feedback control for the augmented system is u (t ) = −kx (t ) − kq q (t )

(3.1.85)

Substituting (3.1.85) into (3.1.79), ⎡ dx ⎤ ⎢ dt ⎥ ⎡ A − bk ⎢ ⎥= ⎢ ⎢ dq ⎥ ⎢⎣ c ⎢⎣ dt ⎥⎦

−bkq ⎤ ⎡ x (t ) ⎤ ⎡ w ⎤ ⎥⎢ ⎥+⎢ ⎥ 0 ⎥⎦ ⎣ q (t ) ⎦ ⎣ 0 ⎦

(3.1.86)

Assuming that Ca is nonsingular, it is always possible to find k and kq such that eigenvalues of system (3.1.86) are in the left half of the s-plane. Because w is a constant vector, steady state values of x(t) and q(t) will be constants for a stable closed-loop dq = 0. From Equation 3.1.77, lim y(t ) = 0. Finally, t →∞ dt the control law (3.1.85) can be represented as system. Hence, in steady state,

t

∫

u (t ) = −kx (t ) − kq y (t ) dt

(3.1.87)

0

Hence, by feeding back the integral of the output in addition to states, we can have the output go to zero in the presence of a constant disturbance vector with unknown magnitudes (Kailath, 1980).


94


EXAMPLE 3.4: INTEGRAL OUTPUT FEEDBACK For the spring-mass-damper system (Chapter 2, Figure 2.4), state equations are defined by (2.5.16) and (2.5.17). In addition, consider a constant but unknown disturbance force acting on the mass. Case I: Position Output If the position of the mass, x, is measured by a sensor, y (t ) = cx (t )

(3.1.88)

c = [1

(3.1.89)

where 0]

In this case, ⎡ 0 b⎤ ⎢ ⎥ = rank ⎢ −α / m 0⎦ ⎢ 1 ⎣

⎡A rank ⎢ a ⎣c

1 −β / m 0

0⎤ ⎥ 1⎥ = 3 0 ⎥⎦

(3.1.90)

Therefore, the condition (3.1.81) is satisfied and the augmented system is controllable. For the control law (3.1.52), define k a = [k

kq ] = [ k1

k2

kq ]

(3.1.91)

Then, the closed-loop characteristic polynomial is ⎞ ⎛β ⎛α ⎞ det( sI − ( Aa − b a k a )) = s 3 + ⎜ + k2 ⎟ s 2 + ⎜ + k1 ⎟ s + kq ⎠ ⎝m ⎠ ⎝m

(3.1.92)

Let the desired characteristic polynomial be det(sI − ( Aa − b a k a )) = s 3 + α1s 2 + α 2 s + α 3

(3.1.93)

Matching the coefficients of polynomials,

k1 = α 2 −

β ; m

k2 = α1 −

α ; m

and

kq = α 3

(3.1.94)



95

Case II: Velocity Output If the velocity of the mass, x , is measured by a sensor, y (t ) = cx (t )

(3.1.95)

c = [0

(3.1.96)

where 1]

In this case,

⎡A rank ⎢ a ⎣c

⎡ 0 b⎤ ⎢ rank = ⎥ ⎢ −α / m 0⎦ ⎢ 0 ⎣

1 −β / m 1

0⎤ ⎥ 1⎥ = 2 0 ⎥⎦

(3.1.97)

Therefore, the condition (3.1.48) is not satisfied, and the augmented system is not controllable.

3.2 COMPUTATION OF STATE FEEDBACK GAIN MATRIX FOR A MULTIINPUT SYSTEM The state feedback control law is u(s ) = r(s ) − Kx (s )

(3.2.1)

where r(s) and K are reference input vector and state feedback gain matrix, respectively. Using (2.13.8) and (2.13.18), u( s ) = r( s ) − K Ψ ( s )ξ( s )

(3.2.2)

Substituting (3.2.2) into (2.13.6),

[ DhcS(s ) + ( Dlc + K )Ψ (s )] ξ(s ) = r(s )

(3.2.3)

ξ( s ) = Dk−1 ( s )r( s )

(3.2.4)

Equation 3.2.3 yields

where


96


Dk (s ) = Dhc S (s ) + ( Dlc + K )Ψ (s ) = D (s ) + K Ψ (s )

(3.2.5)

Substituting (3.2.4) into (2.13.5), y (s ) = GK (s )r(s )

(3.2.6)

GK (s ) = N (s ) DK−1 (s )

(3.2.7)

where

GK (s ) is the closed-loop transfer function with full state feedback. Examining the structures of (3.2.5) and (3.2.7), the following points (Kailath, 1980) should be noted: • •

State feedback does not alter the numerator polynomial N(s). State feedback does not alter Dhc . Therefore, the state feedback control cannot change the column degrees of D(s).

Let the characteristic polynomial of the closed-loop system be described by the following monic polynomial: α( s ) = s n + α1s n−1 + … + α n−1s + α n

(3.2.8)

Assume that the column degrees (Chapter 2, Section 2.8) of D(s) are arranged as m

k1 ≤ k2 ≤ … ≤ km ;

∑k = n i

(3.2.9)

i =1

Rewrite the characteristic polynomial (3.2.8) as follows: α( s ) = s n + α1 ( s )s n−k1 + α 2 ( s )s n−k1−k2 + … + α m ( s )

(3.2.10)

where α i (s ) is a polynomial of degree less than ki . Then, it can be verified that ⎡s k1 ⎢ ⎢ det ⎢ ⎢ ⎢ ⎢ ⎣

+ α1 (s ) −1 . 0 0

α 2 (s ) s k2 . 0 0

. . . . .

0 0 . s

km −1

−1

α m (s ) ⎤ ⎥ 0 ⎥ . ⎥ = α(s ) ⎥ 0 ⎥ s km ⎥⎦

(3.2.11)



97

From (3.2.7), the closed-loop characteristic polynomial is given by det( DK (s )) . Because α(s ) is a monic polynomial, −1 det(sI − Ac + Bc K ) = det( Dhc DK ) = α(s )

(3.2.12)

−1 det( S (s ) + Dhc ( Dlc + K )Ψ (s )) = α(s )

(3.2.13)

From (3.2.5),

Comparing (3.2.11) and (3.2.13), ⎡s k1 ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣

+ α1 (s ) −1 . 0 0

α 2 (s ) s k2 . 0 0

. . . . .

0 0 . s km −1 −1

α m (s ) ⎤ ⎥ 0 ⎥ −1 . ⎥ = S (s ) + Dhc ( Dlc + K )Ψ (s ) (3.2.14) ⎥ 0 ⎥ s km ⎥⎦

Using the definition of S (s ) , Equation 3.2.14 yields ⎡α1 (s ) ⎢ ⎢ −1 K Ψ(s ) = Dhc ⎢ . ⎢ ⎢ 0 ⎢ 0 ⎣

α 2 (s ) 0 . 0 0

. . . . .

0 0 . 0 −1

α m (s ) ⎤ ⎥ 0 ⎥ . ⎥ − Dlc Ψ (s ) ⎥ 0 ⎥ 0 ⎥⎦

(3.2.15)

Factoring out Ψ(s ) on the right-hand side yields the state feedback gain matrix K. Equation 3.2.11 is not a unique way to express the desired characteristic polynomial α(s ). As shown in Example 3.5, there are other matrices with their determinants equal to α(s ). Therefore, the state feedback gain matrix K is not unique to obtain the desired closed-loop characteristic polynomial α(s ) for a multiinput system (Kailath, 1980). EXAMPLE 3.5 For DH (s ) in Equation 2.13.32, k1 = 2 and k2 = 1. And, ⎡2 / 3 Dhc = ⎢ ⎣ 0

−2 / 3 ⎤ ⎥ 1 ⎦

and

Let the desired characteristic polynomial be

⎡16 / 3 Dlc = ⎢ ⎣ 0

10 0

−2 ⎤ ⎥ 4⎦

(3.2.16)


98


α(s ) = s 3 + 30 s 2 + 300 s + 1000

(3.2.17)

Because k1 = 2 and k2 = 3, this polynomial can be rearranged to be α(s ) = s 3 + α1 (s )s + α 2 (s )

(3.2.18)

where α1 (s ) = 30 s + 300

α 2 (s ) = 1000

and

(3.2.19)

From (3.2.15), ⎡α (s ) ( K c + Dlc )Ψ(s ) = Dhc ⎢ 1 ⎣ −1

α 2 (s ) ⎤ ⎥ 0 ⎦

(3.2.20)

where ⎡s ⎢ Ψ(s ) = ⎢ 1 ⎢0 ⎣

0⎤ ⎥ 0⎥ 1 ⎥⎦

(3.2.21)

From (3.2.20), ⎡30 ( K c + Dlc )Ψ(s ) = Dhc ⎢ ⎣0

300 −1

1000 ⎤ ⎥ Ψ (s ) 0 ⎦

(3.2.22)

300 −1

1000 ⎤ ⎥ 0 ⎦

(3.2.23)

2006 / 3 ⎤ ⎥ −4 ⎦

(3.2.24)

Equating coefficients of Ψ(s ) on both sides, ⎡30 K c + Dlc = Dhc ⎢ ⎣0 Therefore, ⎡ 44 / 3 Kc = ⎢ ⎣ 0 Alternatively,

572 / 3 −1



99

⎡α (s ) ( K c + Dlc )Ψ (s ) = Dhc ⎢ 1 ⎣ −10

α 2 (s ) / 10 ⎤ ⎥ 0 ⎦

(3.2.25)

Using (3.2.19), ⎡30 ( K c + Dlc )Ψ (s ) = Dhc ⎢ ⎣0

300 −10

100 ⎤ ⎥ Ψ (s ) 0 ⎦

(3.2.26)

300 −10

100 ⎤ ⎥ 0 ⎦

(3.2.27)

Equating coefficients of Ψ(s) on both sides, ⎡30 K c + Dlc = Dhc ⎢ ⎣0 Therefore, ⎡ 44 / 3 Kc = ⎢ ⎣ 0

50 / 3 −10

206 / 3 ⎤ ⎥ −4 ⎦

(3.2.28)

There exists another K c matrix to achieve the desired characteristic polynomial (3.2.18).

3.3 STATE FEEDBACK GAIN MATRIX FOR A MULTIINPUT SYSTEM FOR DESIRED EIGENVALUES AND EIGENVECTORS The closed-loop system is described by GK (s ) = N (s ) DK−1 (s )

(3.3.1)

Dk (s ) = Dhc S (s ) + ( Dlc + K )Ψ (s ) = D (s ) + K Ψ (s )

(3.3.2)

D (s ) = Dhc S (s ) + Dlc Ψ (s )

(3.3.3)

where

and

Let the desired eigenvalues of the closed-loop system be μ i ; i = 1, 2,..., n . In this case,


100


det Dk (μ i ) = 0; i = 1, 2,...., n

(3.3.4)

It can be shown that eigenvectors fi associated with the system matrix of the controller form realization (Chapter 2, Section 2.13) of the transfer matrix GK (s ) can be expressed as fi = Ψ(μ i )p i

(3.3.5)

DK (μ i )p i = 0

(3.3.6)

D (μ i )p i + K Ψ (μ i )p i = 0

(3.3.7)

where the vector p i satisfies

Using (3.3.2) and (3.3.6),

Substituting (3.3.5) into (3.3.7), Kfi = −g i ;

i = 1, 2, …, n

(3.3.8)

where g i = D (μ i )p i

(3.3.9)

Equation 3.3.8 can be put in the following form (Kailath, 1980): K ⎡⎣f1

f2

.

.

f n ⎤⎦ = − ⎡⎣g1

g2

.

.

g n ⎤⎦

(3.3.10)

Assuming that eigenvectors {fi } are linearly independent, K = − ⎡⎣g1

g2

.

.

g n ⎤⎦ ⎡⎣f1

f2

.

.

f n ⎤⎦

−1

(3.3.11)

Linear independence of eigenvectors fi is guaranteed when eigenvalues μ i are distinct. Although it is possible to get a unique solution for K, even when some of the eigenvalues are repeated, it will be assumed that eigenvalues μ i are distinct for the following discussion. To use Equation 3.3.11, g i and fi are obtained (Kailath, 1980) as follows:



101

1. Choose unrepeated closed-loop eigenvalues μ i . 2. Choose the desired closed-loop eigenvectors fi to satisfy the following constraints: a. With reference to Equation 3.3.5, fi must belong to the range space of Ψ(μ i ) . b. All the eigenvectors fi are linearly independent. c. If fi corresponds to a complex eigenvalue μ i , the complex conjugate of fi must correspond to the eigenvalue which is the complex conjugate of μ i . 3. Use Equation 3.3.5 to solve for p i . Premultiply both sides of Equation 3.3.5 by Ψ T (μ i ): Ψ T (μ i )Ψ (μ i )p i = Ψ T (μ i )fi

(3.3.12)

Because the matrix Ψ(μ i ) is of full rank, the square matrix Ψ T (μ i )Ψ (μ i ) is nonsingular. Therefore, p i = (Ψ T (μ i )Ψ(μ i ))−1 Ψ T (μ i )fi

(3.3.13)

4. From Equation 3.3.9, g i is computed. EXAMPLE 3.6: ELECTRONICS NAVIGATION OR GYRO BOX (SCHULTZ AND INMAN, 1994) An electronics navigation or gyro box is mounted on passive spring-damper isolators located at the bottom four corners (Figure 3.4). There are isolators along x and z directions at each bottom four corner. There are also three actuators providing active control forces u1 , u2, and u3 as shown in Figure 3.4. Free body diagram of the system is shown in Figure 3.5. Applying Newton’s law, the system of differential equations of motion is written as + Eq + K s q(t ) = v (t ) Mq

(3.3.14)

v (t ) = B f u(t )

(3.3.15)

where the mass matrix M, the damping matrix E, the stiffness matrix K s , and the input force matrix B f are expressed as: ⎡m ⎢ M = ⎢0 ⎢0 ⎣

0 m 0

0⎤ ⎥ 0⎥ I yy ⎥⎦

⎡ 4c z ⎢ E=⎢ 0 ⎢ 0 ⎣

0 4c x −4 z c x

⎤ 0 ⎥ −4 z c x ⎥ 4(c z 2x + c x 2z ) ⎥⎦

(3.3.16a,b)


102


x

x

qz

cg

qx

z

cx

cx

u2 kx

kx u1 cz

u3

kz

cz

kz

FIGURE 3.4 Electronics navigation or gyro box.

x

x

qz

cg

qx

z

u1 k x (qx

z ) c x (q x

k z (qz

u3

z ) x )

k x (q x

u2 cz (q z

x )

FIGURE 3.5 Free body diagram of gyro box.

k z (qz

x )

cz (q z

z ) x )

c x (q x

z )



⎡ 4 kz ⎢ Ks = ⎢ 0 ⎢ 0 ⎣

0 4kx −4 z k x

⎤ 0 ⎥ −4 z k x ⎥ 4( kz 2x + k x 2z ) ⎥⎦

103

⎡1 ⎢ Bf = ⎢ 0 ⎢ x ⎣

0 −1 z

1 ⎤ ⎥ 0 ⎥ − x ⎥⎦

(3.3.17a,b)

For a mechanical system, it is usual to define states as p1 = q

and

p 2 = q

(3.3.18a,b)

Then, p 1 = p 2

(3.3.19)

p 2 = − M −1K s p1 − M −1Ep 2 + M −1v (t )

(3.3.20)

and from (3.3.14),

Putting (3.3.19) and (3.3.20) in the matrix form, p = Ap p(t ) + Bp v (t )

(3.3.21)

where ⎡p ⎤ p = ⎢ 1⎥ ; ⎣p 2 ⎦

⎡ 0 Ap = ⎢ −1 ⎣− M Ks

I ⎤ ⎥; − M −1E ⎦

⎡ 0 ⎤ Bp = ⎢ −1 ⎥ ⎣M ⎦

(3.3.22a,b,c)

Expressing (3.3.14) in the matrix fraction description (MFD) form, q(s ) = N (s ) D −1 (s )v (s )

(3.3.23)

where N (s ) = I

and

D (s ) = ( Ms 2 + Es + K s )

(3.3.24a,b)

Here, ⎡s 2 ⎢ D (s ) = Dhc ⎢ 0 ⎢0 ⎣

0 s2 0

0⎤ ⎥ 0 ⎥ + Dc Ψ (s ) s 2 ⎥⎦

(3.3.25)


104


where Dhc = M

(3.3.26)

Dc = ⎡ 4 cz ⎢ ⎢ 0 ⎢⎣ 0

4 kz 0 0

0 4 cx −4 z cx

0 4 kx −4 z k x

0 −4 z cx 4 (cz 2x + cx 2z )

⎤ 0 ⎥ −4 z k x ⎥ 4 ( kz 2x + k x 2z )⎥⎦

(3.3.27)

and ⎡s ⎢ ⎢1 ⎢0 Ψ(s ) = ⎢ ⎢0 ⎢0 ⎢ ⎢⎣0

0⎤ ⎥ 0⎥ 0⎥ ⎥ 0⎥ s⎥ ⎥ 1 ⎥⎦

0 0 s 1 0 0

(3.3.28)

Let the desired characteristic polynomial for the closed-loop system be α(s ) = s 6 + β1s 5 + β2 s 4 + β3 s 3 + β 4 s 2 + β5 s + β6

(3.3.29)

Here, k1 = k2 = k3 = 2 . Therefore, Equation 3.3.29 should be expressed as α(s ) = s 6 + s 4 α1 (s ) + s 2 α 2 (s ) + α 3 (s )

(3.3.30)

where α1 (s ) = β1s + β2 ,

α 2 (s ) = β3 s + β 4 ,

and

α 3 (s ) = β5 s + β6

(3.3.31)

It can be easily verified that ⎡ 2 ⎢s + α1 (s ) ⎢ α(s ) = det ⎢ − γ ⎢ 0 ⎢ ⎢ ⎣

α 2 (s ) γ s2 −λ

α 3 (s ) ⎤ γλ ⎥⎥ 0 ⎥ ⎥ s2 ⎥ ⎥ ⎦

(3.3.32)



105

where γ and λ are any arbitrary real numbers. Factoring out Ψ(s ) on both sides of Equation 3.2.15, K= ⎡β1 ⎢ Dhc ⎢ 0 ⎢⎣ 0

β2 −γ 0

β3 / γ 0 0

β4 / γ 0 −λ

β5 / ( γλ ) 0 0

β6 / ( γλ )⎤ ⎥ 0 ⎥ − Dc 0 ⎥⎦

(3.3.33)

This result clearly indicates the gain matrix K is not unique because γ and λ can be chosen arbitrarily. Furthermore, this gain matrix is associated with the controller form state space realization (Chapter 2, Section 2.13), which is developed below. Following definitions (2.13.16) and (2.13.17), ⎡0 ⎢ ⎢1 ⎢0 Ac0 = ⎢ ⎢0 ⎢0 ⎢ ⎢⎣0

0 0 0 0 0 0

0 0 0 1 0 0

0 0 0 0 0 0

0 0 0 0 0 1

0⎤ ⎥ 0⎥ 0⎥ ⎥, 0⎥ 0⎥ ⎥ 0 ⎥⎦

⎡1 ⎢ ⎢0 ⎢0 Bc0 = ⎢ ⎢0 ⎢0 ⎢ ⎢⎣0

0 0 1 0 0 0

0⎤ ⎥ 0⎥ 0⎥ ⎥ 0⎥ 1⎥ ⎥ 0 ⎥⎦

(3.3.34)

Therefore, −1 Ac = Ac0 − Bc0 Dhc Dlc =

⎡ 4 cz ⎢ ⎢ m ⎢ 1 ⎢ ⎢ 0 ⎢ ⎢ 0 ⎢ ⎢ ⎢ 0 ⎢ ⎢⎣ 0

4 kz m 0

0

0

0

0 4 cx m 1

0

0

4 kx m 0

−4 z cx m 0

0

−4 z cx I yy

−4 z k x I yy

4 (cz 2x + cx 2z ) I yy

0

0

0

1

0 0

⎤ ⎥ ⎥ 0 ⎥ ⎥ −4 z k x ⎥ m ⎥ ⎥ 0 ⎥ 2 2 4 ( kz x + k x z ) ⎥ ⎥ I yy ⎥ ⎥⎦ 0 0

(3.3.35)


106


⎡1 / m ⎢ ⎢0 ⎢0 −1 Bc = Bc0 Dhc =⎢ ⎢0 ⎢0 ⎢ ⎣0

0 0 1/ m 0 0 0

⎤ 0 ⎥ 0 ⎥ ⎥ 0 ⎥ 0 ⎥ 1 / I yy ⎥ ⎥ 0 ⎦

(3.3.36)

The state equations are x = Ac x (t ) + Bc u(t )

(3.3.37)

u(t ) = − Kx (t )

(3.3.38)

with the control law

where K is given by Equation 3.3.33. The control law is not directly implementable because sensors provide the states p(t), not x(t). Therefore, to implement this controller, it is necessary to find the similarity transformation that will convert the state space realization (3.3.21) to the state space realization (3.3.37). This is where the similarity transformation (2.13.46) can be used for parameter values as follows. z = 0.11 m, kz = 17500 N/m, c z = 0.002 kz x = 0.08 m, k x = 8750 N/m, c x = 0.002 k x m = 10 kg, I yy = 0.0487 kg-m-m Then ⎡ −14 ⎢ ⎢ 1 ⎢ 0 Ac = ⎢ ⎢ 0 ⎢ 0 ⎢ ⎢⎣ 0

700 0 0 0 0 0

0 0 −7 1 158.11 0

0 0 −350 0 79055.44 0

0 0 0.77 0 −35.79 1

⎤ 0 ⎥ 0 ⎥ ⎥ 385 ⎥ 0 ⎥ −17895.277 ⎥ ⎥ 0 ⎥⎦

(3.3.39)



107

Define p(t ) = Tz (t )

(3.3.40)

where the similarity transformation matrix T is defined by Equation 2.13.46. Substituting (3.3.40) into (3.3.21), z = Acc z (t ) + Bcc v (t )

(3.3.41)

where Acc = T −1ApT = ⎡−14 ⎢ ⎢ 1 ⎢ 0 ⎢ ⎢ 0 ⎢ 0 ⎢ ⎣ 0

700 0 0 0 0 0

0 0 −7 1 0.77 0

0 0 −350 0 385 0

0 0 158.11 0 −35.79 1

⎤ 0 ⎥ 0 ⎥ 79055.44 ⎥ ⎥ 0 ⎥ −17895.277⎥ ⎥ 0 ⎦

(3.3.42)

and Bcc = T −1Bp = Bc0

(3.3.43)

It is seen that structures of Acc and Bcc are similar to those of Ac and Bc , respectively. Using (2.13.21) and (2.3.22), new Dhc = Dhcc and new Dc = Dcc are obtained as follows: Dhcc = I 6

(3.3.44)

and ⎡14 ⎢ Dcc = ⎢ 0 ⎢0 ⎣

−700 0 0

0 7 −0.77

0 350 −385

0 −158.11 35.79

⎤ 0 ⎥ −79055.44 ⎥ (3.3.45) 17895.277 ⎥⎦

Now, formulae (3.2.15) and (3.3.11) should be used with these new Dhc = Dhcc and new Dc = Dcc. A Matlab code 3.1 is attached to do the following:


108


1. Locate the eigenvalues of the closed-loop system, μ i , such that the damping factor of each vibratory mode is five times the corresponding value of the open-loop system, and undamped natural frequencies remain unchanged. 2. Locate the eigenvalues of the closed-loop system as described in part (1), and eigenvectors as: ⎡ μ1 ⎤ ⎢ ⎥ ⎢1⎥ ⎢0⎥ f1 = ⎢ ⎥ ; ⎢0⎥ ⎢0⎥ ⎢ ⎥ ⎢⎣ 0 ⎥⎦

⎡μ 2 ⎤ ⎢ ⎥ ⎢1⎥ ⎢0⎥ f2 = ⎢ ⎥ ; ⎢0⎥ ⎢0⎥ ⎢ ⎥ ⎢⎣ 0 ⎥⎦

⎡0⎤ ⎢ ⎥ ⎢0⎥ ⎢μ ⎥ f3 = ⎢ 3 ⎥ ; ⎢1⎥ ⎢0⎥ ⎢ ⎥ ⎢⎣ 0 ⎥⎦

⎡0⎤ ⎢ ⎥ ⎢0⎥ ⎢μ ⎥ f4 = ⎢ 4 ⎥ ; ⎢1⎥ ⎢0⎥ ⎢ ⎥ ⎢⎣ 0 ⎥⎦

⎡0⎤ ⎢ ⎥ ⎢0⎥ ⎢0⎥ f5 = ⎢ ⎥ ; ⎢0⎥ ⎢μ ⎥ ⎢ 5⎥ ⎢⎣ 1 ⎥⎦

⎡0⎤ ⎢ ⎥ ⎢0⎥ ⎢0⎥ f6 = ⎢ ⎥ ⎢0⎥ ⎢μ ⎥ ⎢ 6⎥ ⎢⎣ 1 ⎥⎦

(3.3.46)

It should be noted the choice of eigenvectors (3.3.46) satisfies the required constraints. MATLAB PROGRAM 3.1: FULL STATE FEEDBACK CONTROL NAVIGATION OR GYRO BOX % clear all close all % lz=0.11; m=10; kz=17500; Iyy=0.0487; lx=0.08; kx=8750; %

OF

ELECTRONICS



M=diag([m m Iyy]); Ks=[4*kz 0 0;0 4*kx -4*lz*kx;0, -4*lz*kx 4*((kx*lz*lz)+(kz*lx*lx))]; E=0.002*Ks; Bf=[1 0 1;0 -1 0;lx lz -lz]; % Ac0=zeros(6,6); Ac0(2,1)=1.; Ac0(4,3)=1; Ac0(6,5)=1; % Bc0=zeros(6,3); Bc0(1,1)=1; Bc0(3,2)=1; Bc0(5,3)=1; % Dhc=M; Dlc=[E(1,1) Ks(1,1) 0 0 0 0;0 0 E(2,2) Ks(2,2) E(2,3) Ks(2,3);0 0 E(3,2) Ks(3,2) E(3,3) Ks(3,3)]; Ac=Ac0-Bc0*inv(Dhc)*Dlc; Bc=Bc0*inv(Dhc); % Ap=[0*eye(3) eye(3);-inv(M)*Ks -inv(M)*E]; eigop=eig(Ap); Bp=[0*eye(3);inv(M)]; CI=[Bp Ap*Bp]; %Soloution of Eq. (2.13.39) Coeff=inv(CI)*Ap*Ap*Bp; Beta=Coeff(4:6,:); % %Similarity Transformation Matrix T, Eq. (2.13.46)

109


110


for i=1:3 T(:,2*i-1)=Bp(:,i); T(:,2*i)=Ap*Bp(:,i)-Bp*Beta(:,i); end % Acc=inv(T)*Ap*T; Bcc=inv(T)*Bp; % %New Dhc=Dhcc and New Dlc=Dlcc % Accd=Acc-Ac0; Dhcc=eye(3); Dlcc=-[Accd(1,:);Accd(3,:);Accd(5,:)]; % %Find Closed-Loop Eigenvalues % im=sqrt(-1); for i=1:6 rp=real(eigop(i)); if (imag(eigop(i))0.)imm=im; end %Open Loop Frequencies and Damping Factors ogg(i)=abs(eigop(i)); zeta(i)=-rp/ogg(i); %Closed Loop Damping Factor=5* Open Loop Damping Factor zetacl(i)=5*zeta(i);



111

%Desired Closed Loop Eigenvalues mu(i)=-zetacl(i)*ogg(i)+imm*ogg(i)*sqrt(1.zetacl(i)^2); end % % Desired Closed-Loop Characteristic Polynomial % chcl=poly(mu); % %Non-unique State Feedback Gain Matrix KU for %Specified Eigenvalues %gamma and lambda can be chosen arbitrarily. % gamma=10; lambda=1000; % ve1=[chcl(2) chcl(3)]; ve2=[chcl(4) chcl(5)]/gamma; ve3=[chcl(6) chcl(7)]/(gamma*lambda); % mave=[ve1 ve2 ve3;0 -gamma 0 0 0 0;0 0 0 -lambda 0 0]; % KU=Dhcc*mave-Dlcc; % % Unique State Feedback Gain Matrix KK for Specified %Eigenvalues and Eigenvectors %Choose eigenvectors appropriately ff(:,1)=[mu(1) 1 0 0 0 0].'; ff(:,2)=[mu(2) 1 0 0 0 0].'; ff(:,3)=[0 0 mu(3) 1 0 0].';


112


ff(:,4)=[0 0 mu(4) 1 0 0].'; ff(:,5)=[0 0 0 0 mu(5) 1].'; ff(:,6)=[0 0 0 0 mu(6) 1].'; % for i=1:6 psi=[mu(i) 1 0 0 0 0;0 0 mu(i) 1 0 0;0 0 0 0 mu(i) 1].'; DD=Dhcc*(mu(i)^2)+Dlcc*psi; pp=inv(psi.'*psi)*psi.'*ff(:,i); %Equation (3.3.9) gg(:,i)=DD*pp; end % %Solution of Eq. (3.3.11) % KK=-gg*inv(ff);

3.4 FUNDAMENTALS OF OPTIMAL CONTROL THEORY Consider a general dynamic system of order n, dx = f ( x (t ), u(t )); dt

x (0 ) = x 0

(3.4.1)

where x(t ) and u(t ) are n-dimensional state and m-dimensional input vectors, respectively. The objective is to determine u(t ); 0 ≤ t ≤ tf , such that the following objective function (Ray, 1981) is minimized or maximized: tf

I (u(t )) = G ( x (t f )) +

∫ F (x, u)dt

(3.4.2)

0

Note that state equations serve as constraints for the optimization of I. In addition, constraints on the input of the following types will be considered:



113

ui* ≤ ui ≤ ui*

3.4.1 NECESSARY CONDITIONS

FOR

(3.4.3)

OPTIMALITY

Let u o (t ) be a candidate for the optimal input vector, and let the corresponding state vector be x o (t ) , i.e., dx 0 (t ) = f ( x o (t ), u o (t )) dt

(3.4.4)

In order to see whether u o (t ) is indeed an optimal solution, this candidate optimal input is perturbed (Ray, 1981) by a small amount δu(t ) ; i.e., u(t ) = u o (t ) + δu(t )

(3.4.5)

The change in the value of the objective function can be written as δI = I (uo (t ) + δu(t )) − I (uo (t )) ⎛ ∂G ⎞ = ⎜ ⎟ δx(t f ) + ⎝ ∂x ⎠

tf

∫ 0

(3.4.6) ⎤ ⎡ ⎡⎛ ∂F ⎞ ⎛ ∂G ⎞ ⎛ ∂F ⎞ ⎤ t t dt + F t + δ x δ u ( ) f( ) δ + ⎜ ⎟ ⎜ ⎟ ⎜ ⎟ f ⎥ f ⎢ f ⎥ ⎢ ⎝ ∂x ⎠ ⎝ ∂u ⎠ ⎦ ⎦ ⎣ ⎣⎝ ∂x ⎠

If the solution of (3.4.1) with u(t ) given by (3.4.5) is x o (t ) + δx (t ), d (x o (t ) + δx(t )) = f (x o (t ) + δx(t ), uo (t ) + δu(t )) dt

(3.4.7)

Linearizing (3.4.7), d (δx ) ∂f ∂f = ( )δx (t ) + ( )δu(t ) dt ∂x ∂u

(3.4.8)

Multiplying Equation 3.4.8 by λ T (t ) and integrating from 0 to t f , tf

∫ 0

λ T (t )

d (δx ) dt − dt

tf

⎡⎛

⎞

⎛

⎞

⎤

∫ λ (t )⎢⎣⎜⎝ ∂∂xf ⎟⎠ δx(t ) + ⎜⎝ ∂∂uf ⎟⎠ δu(t )⎥⎦ dt = 0 T

(3.4.9)

0

where λ (t ) is an n-dimensional vector. Adding (3.4.9) to (3.4.6) and evaluating the first integral in (3.4.9) by parts (Ray, 1981),


114


⎤ ⎤ ⎡ ⎡⎛ ∂G ⎞ ⎛ ∂G ⎞ δI = ⎢F (t f ) + ⎜ ⎟ f (t f )⎥ δt f + λ T (0 )δx(0)) + ⎢⎜ ⎟ − λ T (t f )⎥ δx(t f ) ⎝ ∂x ⎠ ⎦ ⎦ ⎣ ⎣⎝ ∂x ⎠ tf

+

∫ 0

⎡⎛ ∂H ⎞ ⎛ ∂H ⎞ dλT ⎤ δx⎥dt ⎟ δx + ⎜ ⎟ δu(t ) + ⎢⎜ ⎝ ∂u ⎠ dt ⎦ ⎣⎝ ∂x ⎠

(3.4.10)

where the function H is known as Hamiltonian, defined as follows: H = F (x, u) + λ T (t )f (x, u)

(3.4.11)

Because λ (t ) is arbitrary, it is chosen to satisfy ⎛ ∂H ⎞ dλT = −⎜ ⎟ ⎝ ∂x ⎠ dt

(3.4.12)

Terms outside the integral in (3.4.10) are known as boundary conditions terms, which are removed for the specified problem. For example, if t f and x(0 ) are specified, δt f = 0 and δx(0 ) = 0 , and the third outside term in (3.4.10) vanishes under the following condition: ⎛ ∂G ⎞ λ T (t f ) = ⎜ ⎟ ⎝ ∂x ⎠ t =t f

(3.4.13)

Hence, the Equation 3.4.10 can be written as tf

δI =

⎛

⎞

∫ ⎜⎝ ∂∂Hu ⎟⎠ δudt

(3.4.14)

0

Minimization of I If u o (t ) is an optimal solution, δI ≥ 0

for any perturbation δu(t )

(3.4.15)

Case I: No Constraint on Input For condition (3.4.15) to be true, ⎛ ∂H ⎞ ⎜ ⎟=0 ⎝ ∂u ⎠

for every t

This is the necessary condition for optimality.

(3.4.16)



115

Case II: With Constraints (3.4.3) on Inputs In view of the condition (3.4.15), If uio = ui* ,

⎛ ∂H ⎞ ⎜ ⎟≤0 ⎝ ∂ui ⎠

(3.4.17)

If uio = ui*,

⎛ ∂H ⎞ ⎜ ⎟≥0 ⎝ ∂ui ⎠

(3.4.18)

Maximization of I If u o (t ) is an optimal solution, δI ≤ 0

for any perturbation δu(t )

(3.4.19)

Case I: No Constraint on Input For condition (3.4.19) to be true, ⎛ ∂H ⎞ ⎜ ⎟ = 0 for every t ⎝ ∂u ⎠

(3.4.20)

This is the necessary condition for optimality and is identical to Equation 3.4.16. Case II: With Constraints (3.4.3) on Inputs In view of the condition (3.4.19), If uio = ui*,

⎛ ∂H ⎞ ⎟≥0 ⎜ ⎝ ∂ui ⎠

(3.4.21)

If uio = ui* ,

⎛ ∂H ⎞ ⎜ ⎟≤0 ⎝ ∂ui ⎠

(3.4.22)

3.4.2 PROPERTIES OF HAMILTONIAN FOR AN AUTONOMOUS SYSTEM For an autonomous system, the function f is not an explicit function of time. Therefore, from Equation 3.4.11, dH ∂H dx ∂H du ∂H d λ = + + dt ∂x dt ∂u dt ∂λ dt But

(3.4.23)


116


∂H = fT ∂λ

(3.4.24)

Substituting (3.4.1) and (3.4.12) into (3.4.23), dH ∂H du = dt ∂u dt

(3.4.25)

∂H = 0 . When there are constraints, inputs ∂u can be either maximum or minimum constant values according to Equation 3.14.17 and Equation 3.4.18 or Equation 3.14.21 and Equation 3.4.22, respectively. In this Without any constraints on inputs,

case,

du = 0 . Therefore, for optimal inputs, dt dH =0 dt

(3.4.26)

In other words, Hamiltonian H is a constant along an optimal trajectory for an autonomous system. Special Case Final time t f not specified; i.e., δt f ≠ 0 . In this case, to remove the corresponding boundary condition term in Equation 3.4.10, ⎛ ∂G ⎞ F (t f ) + ⎜ ⎟ f (t f ) = 0 ⎝ ∂x ⎠

(3.4.27)

When final conditions on states are not specified, the condition (3.4.13) holds, and Equation 3.4.27 reduces to H (t f ) = F (t f ) + λ T (t f )f (t f ) = 0

(3.4.28)

The condition (3.4.28) is valid even when some or all states are specified at the final time. For example, consider that all states are specified at the final time. In this case, with the first-order term in the Taylor series expansion (Ray, 1981), x o (t f ) = x (t f + δt f ) = x (t f ) + x (t f )δt f or

(3.4.29)



117

δx (t f ) = x (t f ) − x o (t f ) = −f (t f )δt f

(3.4.30)

It is interesting to note that δx(t f ) ≠ 0 . Substituting Equation 3.4.30 into (3.4.10) and using (3.4.12), tf

δI = [ F (t f ) + λ (t f )f (t f )]δt f + λ (0 )δx(0 ) + + T

T

⎛

⎞

∫ ⎜⎝ ∂∂Hu ⎟⎠ δu(t )dt

(3.4.31)

0

Therefore, the condition (3.4.28) is again needed to remove boundary condition terms. In summary, H (t f ) = 0 when t f is not specified. Because H is a constant, it is concluded that H(t) = 0 along an optimal trajectory for an autonomous system when t f is not specified. EXAMPLE 3.7: TRAVEL OVER MAXIMUM DISTANCE CONSTRAINT ON FINAL VELOCITY

IN

SPECIFIED TIME

WITHOUT

Consider the simple mass m, which is subjected to force f (t ) . The differential equation of motion is x = u

(3.4.32)

where u (t ) =

f (t ) m

(3.4.33)

Find the optimal control input u (t ) , such that the vehicle covers the maximum distance (Biegler, 1982) in a fixed time t f , subject to the following constraints: a ≤ u (t ) ≤ b

(3.4.34)

Without any loss of generality take x(0 ) = 0 . Assume that the vehicle starts from rest; i.e., 0) = 0 x(

(3.4.35)

Solution Define the following state variables: x1 = x

and

x2 = x

(3.4.36)


118


Hence, state equations are ⎡ x1 ⎤ ⎡ x2 ⎤ ⎢ ⎥= ⎢ ⎥ =f ⎣ x2 ⎦ ⎣ u ⎦

(3.4.37)

The objective in this problem is to maximize I = x1 (t f )

(3.4.38)

Therefore, G = x1 (t )

and

F =0

(3.4.39)

The Hamiltonian H for this problem is H = F + λ T f = [ λ1

⎡ x2 ⎤ λ 2 ] ⎢ ⎥ = λ1 x 2 + λ 2 u ⎣u⎦

(3.4.40)

Adjoint equations are d λ1 ∂H =− =0 dt ∂x1

(3.4.41)

d λ2 ∂H =− = − λ1 dt ∂x2

(3.4.42)

Then, conditions on state variables are as follows: x1 (0 ) = 0 ⇒ δx1 (0 ) = 0

(3.4.43a)

x2 (0 ) = 0 ⇒ δx2 (0 ) = 0

(3.4.43b)

x1 (t f )unspecified ⇒ δx1 (t f ) ≠ 0

(3.4.43c)


(3.4.44d)

Also, because the final time t f is fixed,



119

δt f = 0

(3.4.45)

For the boundary condition terms to be zero, ⎛ ∂G ⎞ λ T (t f ) = ⎜ ⎟ ⎝ ∂x ⎠t =t f

(3.4.46)

⎛ ∂G ⎞ λ1 (t f ) = ⎜ ⎟ = 1 ⎝ ∂x1 ⎠t =t f

(3.4.47a)

⎛ ∂G ⎞ λ 2 (t f ) = ⎜ ⎟ =0 ⎝ ∂x2 ⎠t =t f

(3.4.47b)

Therefore,

Solutions of adjoint equations are λ1 (t ) = c

and

λ 2 (t ) = −ct + d

(3.4.48a,b)

where c and d are constants. To satisfy the final conditions, c =1

d = tf

and

(3.4.49a,b)

In other words, λ1 (t ) = 1

and

λ 2 (t ) = −t + t f

(3.4.50a,b)

Optimality condition: ∂H = λ2 ∂u

(3.4.51)

∂H ∂H is not equal to zero for a finite time interval. = λ 2 is plotted in Figure 3.6. ∂u ∂u This implies that the optimal u cannot take any intermediate value between a and b. Furthermore, control input is

∂H > 0 for 0 ≤ t < tf . Equation 3.4.21 implies that the optimal ∂u


120


H u

tf

0

FIGURE 3.6

tf

t

∂H = λ 2 vs. time t. ∂u

u=b

for 0 ≤ t ≤ t f

(3.4.52)

The optimal strategy is the maximum acceleration, which is employed by a rider out of common sense, when the objective is to cover the maximum distance in a fixed time, without any constraint on the vehicle velocity at the final time. Now, the optimal state trajectory can be easily obtained by solving the state equations with given initial conditions: x2 (t ) = bt

and

x1 (t ) =

1 2 bt 2

(3.4.53)

1 Therefore, the maximum value of the objective function I is bt 2f . Along the 2 optimal trajectory, H = bt + (t f − t )b = bt f

(3.4.54)

As expected, the Hamiltonian H is a constant along the optimal trajectory. EXAMPLE 3.8: TRAVEL OVER MAXIMUM DISTANCE IN SPECIFIED TIME CONSTRAINT THAT FINAL VELOCITY IS EQUAL TO ZERO.

WITH

Consider the system, which is same as that in Example 3.7. Find the optimal control input u (t ) such that the vehicle covers the maximum distance in a fixed time t f and ends at rest (Biegler, 1982) subject to the following constraints: a ≤ u (t ) ≤ b

(3.4.55)

Here, it is also given that a < 0 and b > 0. Solution Define the following state variables: x1 = x

and

x2 = x

(3.4.56)



121

Hence, state equations are ⎡ x1 ⎤ ⎡ x2 ⎤ ⎢ ⎥= ⎢ ⎥ =f ⎣ x2 ⎦ ⎣ u ⎦

(3.4.57)

The objective in this problem is to maximize I = x1 (t f )

(3.4.58)

Therefore, G = x1 (t )

and

F =0

(3.4.59a,b)

The Hamiltonian H for this problem is H = F + λ T f = [ λ1

⎡ x2 ⎤ λ 2 ] ⎢ ⎥ = λ1 x 2 + λ 2 u ⎣u⎦

(3.4.60)

Adjoint equations are ∂H d λ1 =− =0 ∂x1 dt

(3.4.61)

∂H d λ2 =− = − λ1 ∂x2 dt

(3.4.62)

Then, conditions on state variables are as follows: x1 (0 ) = 0 ⇒ δx1 (0 ) = 0

(3.4.63a)

x2 (0 ) = 0 ⇒ δx2 (0 ) = 0

(3.4.63b)


(3.4.63c)

x2 (t f ) = 0 ⇒ δx2 (t f ) = 0

(3.4.63d)

Also, because the final time t f is fixed,


122


δt f = 0

(3.4.64)

Now, consider the following boundary condition term: ⎤ ⎡⎛ ∂G ⎞ T ⎢⎜ ⎟ − λ (t f )⎥ δx(t f ) = ⎦ ⎣⎝ ∂x ⎠ ⎡⎛ ∂G ⎞ ⎤ ⎡⎛ ∂G ⎞ ⎤ ⎢⎜ ⎟ − λ1 (t f )⎥ δx1 (t f ) + ⎢⎜ ⎟ − λ 2 (t f )⎥ δx2 (t f ) ⎣⎝ ∂x1 ⎠ ⎦ ⎣⎝ ∂x2 ⎠ ⎦

(3.4.65)

Because δx2 (t f ) = 0 , ⎡⎛ ∂G ⎞ ⎤ ⎤ ⎡⎛ ∂G ⎞ T ⎢⎜ ⎟ − λ (t f )⎥ δx(t f ) = ⎢⎜ ⎟ − λ1 (t f )⎥ δx1 (t f ) ⎦ ⎣⎝ ∂x ⎠ ⎣⎝ ∂x1 ⎠ ⎦

(3.4.66)

Hence, for all boundary condition terms to be zero, ⎛ ∂G ⎞ λ1 (t f ) = ⎜ ⎟ = 1 ⎝ ∂x1 ⎠t =t f

(3.4.67)

Solutions of adjoint equations are λ1 (t ) = c

and

λ 2 (t ) = −ct + d

(3.4.68)

where c and d are constants. To satisfy the final condition, c =1

(3.4.69)

In other words, λ1 (t ) = 1

and

λ 2 (t ) = −t + d

(3.4.70a,b)

Optimality condition: ∂H = λ 2 = −t + d ∂u

(3.4.71)



123

2

d

0

d

0

0

tf

d

t

0

∂H = λ 2 vs. time t. FIGURE 3.7 ∂u

The value of d is not known. Any of these three possibilities exists: d < 0, d = 0, or d > 0.

∂H = λ 2 is plotted in Figure 3.7 for all these possibilities. The first thing to ∂u

∂H is not equal to zero for a finite time interval. This implies that the ∂u optimal u(t) cannot take any intermediate value between a and b. Secondly, d < 0 and d = 0 are ruled out because they will lead to u(t) = a < 0 (Equation 3.4.22). Hence, d > 0 and the optimal control input is note is that

⎧b u (t ) = ⎨ ⎪⎩ a

for for

t ≤ ts t > ts

(3.4.72)

where ts is the switching instant, which is, when λ 2 (t ) changes sign in Figure 3.7, d > 0 . Solving the second state equation, x2 (ts ) = bts

and

x2 (t f ) = bts + a (t f − ts )

(3.4.73)

−a tf b−a

(3.4.74)

For x2 (t f ) = 0 , ts =

3.5 LINEAR QUADRATIC REGULATOR (LQR) PROBLEM Consider the linear system dx = Ax (t ) + Bu(t ); dt

x (0 ) = x 0

(3.5.1)


124


The objective is to drive the state vector x(t ) to the origin of the state space (zero state vector) from any nonzero initial values of states. If a state feedback control law is used, x(t ) will quickly die out provided closed-loop poles are located far inside the left half of the s-plane. However, elements of the feedback gain vector can be large in magnitudes and the control cost can be high. On the other hand, if closed-loop poles are located close to the open-loop poles, there will not be much increase in the rate of decay of x(t ) and a relatively small amount of control action will be required. Hence, the location of closed-loop poles is a trade-off between the rate of decay of x(t ) and the magnitude of control input. To make this trade-off, the following objective function is chosen: tf

∫

1 1 I = x T (t f ) S f x (t f ) + [ x T (t )Qx (t ) + u T (t ) Ru(t )]dt 2 2

(3.5.2)

0

where the final time t f is fixed. Without any loss of generality, matrices S f , Q , and R are chosen to be symmetric (Appendix B). In addition, S f and Q are chosen to be positive semidefinite, and R to be positive definite. Symbolically, these are expressed as S f = S Tf ≥ 0 ,

Q = QT ≥ 0 ,

and

R = RT > 0

Problem Find u(t ); 0 ≤ t ≤ t f , such that the objective function (3.5.2) is minimized.

3.5.1 SOLUTION (OPEN-LOOP OPTIMAL CONTROL) For this optimal control problem, the Hamiltonian (3.4.11) is H=

1 T (x (t )Qx(t ) + uT (t )Ru(t )) + λ T (t )( Ax(t ) + Bu(t )) 2

(3.5.3)

The necessary condition (3.4.16) for optimality yields ∂H = Ru(t ) + BT λ (t ) = 0 ∂u or u(t ) = − R−1BT λ (t )

(3.5.4)



125

The dynamics of λ(t ) is given by Equation 3.4.12 with final conditions (3.4.13). Hence, dλ = −Qx(t ) − AT λ (t ); dt

λ (t f ) = S f x(t f )

(3.5.5)

Equation 3.5.1 and Equation 3.5.5 represent a two-point boundary value problem (TPBVP) which can be solved to find λ (t ) and x(t ). Putting (3.5.1) and (3.5.5) in the matrix form, ⎡ x ⎤ ⎡x⎤ ⎢ ⎥= M ⎢ ⎥ ⎣λ ⎦ ⎣λ ⎦

(3.5.6)

where ⎡ A M=⎢ ⎣ −Q

− BR −1BT ⎤ ⎥ − AT ⎦

(3.5.7)

Solving (3.5.6), ⎡ x(t )⎤ Mt ⎡ x(0 )⎤ ⎢ ⎥= e ⎢ ⎥ ⎣λ (t )⎦ ⎣λ (0 )⎦

(3.5.8)

To determine λ (0 ) , the matrix e Mt is partitioned as follows: ⎡ Ε (t ) e Mt = ⎢ 11 ⎣Ε 21 (t )

Ε11 (t ) ⎤ ⎥ Ε 22 (t ) ⎦

(3.5.9)

From (3.5.8) and (3.5.9), x (t ) = Ε11 (t ) x (0 ) + Ε12 (t )λ(0 )

(3.5.10)

λ(t ) = Ε 21 (t ) x (0 ) + Ε 22 (t )λ(0 )

(3.5.11)

Imposing the condition λ (t f ) = S f x(t f ) , Ε 21 (t f )x(0 ) + Ε 22 (t f )λ (0 ) = S f [Ε11 (t f )x(0 ) + Ε12 (t f )λ (0 )]

(3.5.12)


126


From (3.5.12), λ (0 ) = [Ε 22 (t f ) − S f Ε12 (t f )]−1[ S f Ε11 (t f ) − Ε 21 (t f )]x(0 )

(3.5.13)

Substituting this λ (0 ) into (3.5.8), λ (t ) is obtained. Then, the optimal u(t ) is found from (3.5.4). However, this TPBVP must be solved again if initial conditions change. Furthermore, the control inputs are implementable in an open-loop fashion only, as they are not in the forms of functions of states.

3.5.2 SOLUTION (CLOSED-LOOP OPTIMAL CONTROL) Use the following transformation (Ray, 1981): λ (t ) = S (t )x(t )

(3.5.14)

where S (t ) is a symmetric n × n matrix. Substituting (3.5.14) into (3.5.5), dS dx x (t ) + S (t ) = −Qx (t ) − AT S (t ) x (t ) dt dt

(3.5.15)

⎛ dS −1 T T ⎞ ⎜ + SA − SBR B S + Q + A S ⎟ x(t ) = 0 ⎝ dt ⎠

(3.5.16)

Using (3.5.1),

Because Equation 3.5.16 is true for all x(t ) , dS + SA − SBR −1BT S + Q + AT S = 0 dt

(3.5.17)

Equation 3.4.13 and Equation 3.5.14 yield S (t f ) = S f

(3.5.17b)

Equation 3.5.17 is known as the Riccati equation. This nonlinear differential equation can be numerically solved backward in time to determine S (t ). From (3.5.4) and (3.5.14), u(t ) = − K (t ) x (t ) where

(3.5.18)



K (t ) = R −1BT S (t )

127

(3.5.19)

The structure of (3.5.18) indicates that the K (t ) is the optimal state feedback gain matrix. Because the solution of S (t ) does not depend on system states, this gain is optimal for all initial conditions on states.

3.5.3 CROSS TERM

IN THE

OBJECTIVE FUNCTION

Consider a more general form of the quadratic objective function (Anderson and Moore, 1990): tf

I=

∫

1 T 1 x (t f ) S f x (t f ) + [ x T (t )Qx (t ) + u T (t ) Ru(t ) + 2x T (t ) Nu(t )]dt 2 2

(3.5.20)

0

It can be seen that x T (t )Qx (t ) + 2 x T (t ) Nu(t ) + u T (t ) Ru(t ) = x T (t )Qm x (t ) + v T (t ) Rv (t ) (3.5.21) where Qm = Q − NR −1N T

(3.5.22)

v (t ) = u(t ) + R −1N T x (t )

(3.5.23)

and

Equation 3.5.21 can be proved by simply multiplying out the terms on the righthand side. Hence, Equation 3.5.20 can be rewritten as tf

∫

1 1 I = x T (t f ) S f x (t f ) + [ x T (t )Qm x (t ) + v T (t ) Rv (t )]dt 2 2

(3.5.24)

0

and the plant equation (3.5.1) is modified with Equation 3.5.23: x = Am x (t ) + Bv (t ) where

(3.5.25)


128


Am = A − BR −1N T

(3.5.26)

Assuming that Qm ≥ 0 , Equation 3.5.24 and Equation 3.5.25 constitute a standard LQ problem for which the optimal state feedback control law is v (t ) = − R −1BT Sm (t ) x (t )

(3.5.27)

where Sm = − Sm Am − AmT Sm + Sm BR −1BT Sm − Qm;

Sm (t f ) = S f

(3.5.28)

From (3.5.23) and (3.5.27), u(t ) = − K m (t ) x (t )

(3.5.29)

where the optimal state feedback gain matrix is given by K m (t ) = R −1 ( BT Sm (t ) + N T ) EXAMPLE 3.9: MINIMUM ENERGY CONTROL

OF A

(3.5.30)

DC MOTOR

Consider the position controller shown in Figure 3.8. The actuator is an armaturecontrolled DC motor (Kuo, 1995). The torque produced by the motor Tm (t ) is proportional to the armature current ia (t ) ; i.e., Tm (t ) = ki ia (t )

(3.5.31)

Let the back emf developed across the armature be eb (t ) , i.e., eb (t ) = kb θ m (t )

(3.5.32)

where kb is the back emf constant and θm (t ) is the angular position of the rotor. Applying Kirchoff’s law to the armature, u (t ) = Ra ia (t ) + eb (t )

(3.5.33)

where u (t ) is the input voltage and Ra (t ) is the armature resistance. The inductance of the armature windings has been neglected.



129

Ra o ia u (t )

eb

Tm m

o

FIGURE 3.8 An armature-controlled DC motor.

Applying Newton’s second law, + B θ = T (t ) Jm θ m m m m

(3.5.34)

where Jm and Bm are mass moment of inertia and equivalent viscous damping, respectively. From (3.5.31) to (3.5.33), ki kk u (t ) − i b θ m Ra Ra

Tm (t ) =

(3.5.35)

From (3.5.34) and (3.5.35), + αθ = βu (t ) θ m m

(3.5.36)

where α=

Bm kk + i b J m Ra J m

(3.5.37)

ki Ra J m

(3.5.38)

and β= From (3.5.36), state equations are x = Ax (t ) + bu (t ) where

(3.5.39)


130


⎡θ (t ) ⎤ x(t ) = ⎢ m ⎥ ; ⎣θm (t ) ⎦

⎡0 A=⎢ ⎣0

1 ⎤ ⎥; −α ⎦

⎡0 ⎤ b=⎢ ⎥ ⎣β ⎦

(3.5.40)

Let us define an optimal control problem: Find the input u(t), 0 ≤ t ≤ t f , such that the energy consumed by the DC motor is minimized over a fixed final time t f , with the following initial and final conditions on the states: ⎡0 ⎤ x(0 ) = ⎢ ⎥ ⎣0 ⎦

⎡θ f x(t f ) = ⎢ ⎢⎣ 0

and

⎤ ⎥ ⎥⎦

(3.5.41)

The energy consumed by the DC motor over a fixed final time t f is given by the following integral: tf

tf

∫ T (t)dθ = ∫ T θ dt

Id =

m m

(3.5.42)

(ux2 −kb x22 ) dt

(3.5.43)

m

m

t =0

0

Using (3.5.35), (3.5.40), and (3.5.42) tf

Id =

∫R

ki

0

a

To satisfy the requirements for the existence of solution to the LQ control, the objective function (3.5.43) is modified to be tf

I=

∫R

ki

0

(ux2 −kb x22 ) + γx22 + ρu 2 dt ;

γ >0

and

ρ>0

(3.5.44)

a

Hence, with respect to the definition (3.5.20), tf

1 I= 2 ⎡0 Q = ⎢⎢ 0 ⎢⎣

∫ x Qx + 2 x Nu + u Rudt T

T

T

(3.5.45)

0

⎤ ⎥; kk 2(ρ − i b ) ⎥ Ra ⎥⎦ 0

⎡ 0 N = ⎢⎢ ki ⎢⎣ Ra

⎤ ⎥; ⎥ ⎥⎦

R = 2ρ

(3.5.46)



131

and ⎤ ⎥ 2k k k (2 γ − i b − )⎥ Ra 2 R ρ ⎥⎦

⎡0 ⎢ Qm = ⎢ 0 ⎢ ⎣

0

2 i 2 a

(3.5.47)

For Qm ≥ 0 , ρ and γ must be chosen to satisfy 2γ −

2 ki kb k2 − i2 ≥ 0 Ra 2 Ra ρ

(3.5.48)

In order to have final conditions on states to be zero, the following transformation of the state vector is introduced: x (t ) = x (t ) − x (t f )

(3.5.49)

Then, x (0 ) = − x (t f )

and

x (t f ) = 0

(3.5.50a,b)

Substituting (3.5.49) into (3.5.39), and noting that Ax(t f ) = 0 , x = Ax (t ) + bu (t )

(3.5.51)

Because the objective function does not contain x1 and x2 = x2 , Equation 3.5.45 can be rewritten as tf

1 I= 2

∫ x Qx + 2 x Nu + u Rudt T

T

T

(3.5.52)

0

In the LQ control, there is no constraint on the final conditions. Therefore, the objective function is modified to contain the final state vector:

I=

where

1 T 1 x (t f )S f x (t f ) + 2 2

tf

∫ x Qx + 2x Nu + u Rudt T

0

T

T

(3.5.53)


132


⎡a Sf = ⎢ ⎣0

0⎤ ⎥; b⎦

a > 0 and b > 0

(3.5.54)

Relaxing constraints on final states, the LQ problem can be solved to minimize (3.5.53) for the linear system (3.5.51) with nonzero x (0 ) . By choosing a and b to be large numbers, x (t f ) can be forced to be close to zero.

3.5.4 IMPORTANT CASE: INFINITE FINAL TIME When the final time t f → ∞ , the optimal gain K m (t ) or Sm (t ) turns out to be a constant (Kwakernaak and Sivan, 1972). The system of differential equations (3.5.28) reduces to a system of algebraic equations: Sm Am − Sm BR −1BT Sm + Qm + AmT Sm = 0

(3.5.55)

This is known as the algebraic Riccati equation (ARE) and is probably one of the most commonly used equations in modern control theory. Since the matrix Sm is a constant, the feedback gain matrix also turns out to be a constant as follows: K o = R −1 ( BT Sm + N T )

(3.5.56)

Hence from (3.5.1) and (3.5.56), the closed-loop system dynamics is represented by dx = ( A − BK o ) x (t ) dt

(3.5.57)

The eigenvalues of the matrix A − BK o are the optimal closed-loop poles. In the next section, a method is described to determine these optimal closed-loop poles first via root locus plots. Then, the optimal feedback gain can be obtained via pole placement techniques. Because Qm ≥ 0 , there exists a matrix H known as the square root of Qm , such that Qm = ΗΗ T

(3.5.58)

If ( A, Η ) is observable, Sm is positive definite and is the only solution of ARE, Equation 3.5.55, with this property. Furthermore, the optimal closed-loop system (3.5.57) is asymptotically stable (Anderson and Moore, 1990; Kailath, 1980). EXAMPLE 3.10: A FIRST-ORDER SYSTEM Consider a first-order system



133

x = x + u

(3.5.59)

and the objective function tf

5 1 I = x 2 (t f ) + 2 2

∫ 2x

2

+ 3u 2 dt

(3.5.60)

0

Here, A = 1, b = 1, Q = 2, R = 3, S f = 5

(3.5.61)

Therefore, the Riccati equation becomes S2 S = −2 S + −2 ; 3

S (t f ) = 5

(3.5.62)

Nonlinear differential equation (3.5.62) has to be numerically solved backward in time to determine S(t) for 0 ≤ t ≤ t f . Matlab routine ODE23 or ODE45 can be used for this purpose. Then, the optimal control law will be

u (t ) = −

S (t ) x (t ) ; 3

0 ≤ t ≤ tf

(3.5.63)

If t f → ∞ , ARE becomes

0 = −2 S +

S2 −2 3

(3.5.64)

Solving (3.5.64), S = 3 ± 15

(3.5.65)

In order to have a stable closed-loop system, a “+” sign must be chosen. The optimal state feedback law is u (t ) = − Kx (t ) where

(3.5.66)


134


15 3

K = 1+

(3.5.67)

Substituting (3.5.66) into (3.5.59), the closed-loop system dynamics is 15 x 3

x = −

(3.5.68)

which is stable. EXAMPLE 3.11: ACTIVE SUSPENSION (THOMSON, 1976)

WITH

OPTIMAL LINEAR STATE FEEDBACK

Differential equations of motion for the linear vehicle model shown in Figure 3.9 are m2 x2 = u

(3.5.69)

m1 x1 + λ1 ( x1 − x0 ) = −u

(3.5.70)

x1 = x3

(3.5.71)

x2 = x 4

(3.5.72)

Defining x3 = x1 and x 4 = x2 ,

x2

Body m2 u(t ) Actuator

? u(t ) Wheel

x1

m1 Tire Road

FIGURE 3.9 A quarter car model.

1

x0



135

λ1 u (t ) ( x1 − x0 ) − m1 m1

x3 = −

x 4 =

(3.5.73)

u (t ) m2

(3.5.74)

Assume that the road profile is a step function. In this case, x0 is a constant, and a new set of state variables is defined as xˆ1 = x1 − x0

(3.5.75)

xˆ2 = x2 − x0

(3.5.76)

xˆ3 = xˆ1 = x1

(3.5.77)

xˆ 4 = xˆ2 = x2

(3.5.75)

and the new set of state equations can be written as

⎡ xˆ1 ⎤ ⎡ 0 ⎢ ⎥ ⎢ 0 ⎢ xˆ2 ⎥ ⎢ ⎢ ⎥ = ⎢ − λ1 ⎢ xˆ3 ⎥ ⎢ m1 ⎢ xˆ ⎥ ⎢ ⎣ 4 ⎦ ⎢⎣ 0

0 0

1 0

0

0

0

0

⎡ 0 ⎤ 0⎤ ⎢ ⎥ 0 ⎥ ⎥ ⎡ xˆ1 ⎤ 1⎥ ⎢ ⎥ ⎢ ⎢ 1 ⎥ xˆ ⎥ ⎢ 2 ⎥ + ⎢− ⎥ u (t ) 0 ⎥ ⎢ xˆ3 ⎥ ⎢ m1 ⎥ ⎥⎢ˆ ⎥ ⎢ 1 ⎥ ⎢x ⎥ ⎥ 0 ⎥⎦ ⎣ 4 ⎦ ⎢ ⎢⎣ m2 ⎥⎦

(3.5.76)

The performance index is chosen as

I=

1 2

∞

∫ [ρu + q (x − x ) + q (x − x ) ]dt 2

2

1

0

1

2

2

1

2

(3.5.77)

0

Note the following: ( x0 − x1 ) : tire dynamic deflection

(3.5.78)

( x1 − x2 ): relative wheel travel

(3.5.79)


136


Moreover, the actuator force u(t), which is proportional to the vertical acceleration of the body, is also a measure of the ride discomfort. Hence, the minimization of the objective function I will result in a trade-off between the minimization of the actuator input or ride discomfort and minimization of the weighted sum of tire dynamic deflection and relative wheel travel. Weighting parameters are ρ , q1, and q2 . Because q1 ( x0 − x1 )2 + q2 ( x1 − x2 )2 = q1 xˆ12 + q2 ( xˆ1 − xˆ2 )2

(3.5.80)

= (q1 + q2 )xˆ12 + q2 xˆ22 − 2 q2 xˆ1 xˆ2 the objective function I can be expressed as

1 I= 2

∞

∫ [ρu

2

+ xˆ T Qxˆ ]dt

(3.5.81)

0

where ⎡q1 + q2 ⎢ −q2 Q=⎢ ⎢ 0 ⎢ ⎢⎣ 0

−q2 q2 0 0

0 0 0 0

0⎤ ⎥ 0⎥ 0⎥ ⎥ 0 ⎥⎦

and

⎡ xˆ1 ⎤ ⎢ˆ ⎥ x xˆ = ⎢ 2 ⎥ ⎢ xˆ3 ⎥ ⎢ ⎥ ⎢⎣ xˆ 4 ⎥⎦

(3.5.82a,b)

The optimal control law is u (t ) = −kxˆ = − ⎡⎣ k1

k2

k3

k4 ⎤⎦ xˆ = − k1 xˆ1 − k2 xˆ2 − k3 xˆ3 − k4 xˆ 4

(3.5.83)

where k = ρ−1b T S

(3.5.84)

SA + AT S − Sbρ−1b T S + Q = 0

(3.5.85)

and

Physical realization of the control law:



137

x2

Body m2

Active Feedback x2

Passive Feedback

f (t ) x2 2

x0

2

f (t ) Wheel

x1

m1 Tire

1

Road

x0

FIGURE 3.10 Active or passive vehicle suspension system.

u (t ) = −k1 ( xˆ1 − xˆ2 ) − k3 ( xˆ 3 − xˆ 4 ) − ( k1 + k2 ) xˆ2 − ( k3 + k4 )xˆ 4 = u p + f (t )

(3.5.86)

where u p (t ) = λ 2 ( x1 − x2 ) + β2 ( x1 − x2 )

(3.5.87)

f (t ) = −( k1 + k2 )( x2 − x0 ) − ( k3 + k4 ) x2

(3.5.88)

and

The part of the control input u p (t ) has been applied by a spring and a damper that provide passive feedback of ( x1 − x2 ) and ( x1 − x2 ) (Figure 3.10). The remaining part f(t) is applied by active feedback via sensors to measure ( x2 − x0 ) and x2 , and an actuator to apply the force f(t). In the study by Thompson (1976), the relative distance ( x2 − x0 ) is measured by an ultrasonic transmitter/receiver, and the velocity x2 is obtained by integrating the signal from an accelerometer. The actuator is shown to be an electrohydraulic system.

3.6 SOLUTION OF LQR PROBLEM VIA ROOT LOCUS PLOT: SISO CASE Let the quadratic objective function be


138


1 I= 2

∞

∫ y (t) + ru (t)dt 2

2

(3.6.1)

0

where u (t ) and y (t ) are input and output of a SISO linear system. Note that y 2 (t ) = y T (t ) y (t ) = x T (t )cT cx (t )

(3.6.2)

Referring to (3.5.2), Q = cT c

and

R=r

(3.6.3)

Hence, from Equation 3.5.1, Equation 3.5.4, and Equation 3.5.5, ⎡ dx(t ) ⎤ ⎥ ⎢ ⎡ x(t )⎤ ⎢ dt ⎥ = M ⎢ ⎥ ⎣λ (t )⎦ ⎢ d λ (t ) ⎥ ⎢⎣ dt ⎥⎦

(3.6.4)

where ⎡ A M=⎢ T ⎣ −c c

−r −1bb T ⎤ ⎥ − AT ⎦

(3.6.5)

The matrix M is known as the Hamiltonian matrix. It can be shown (Kwakernaak and Sivan, 1972) that (−1) n Δ(s ) = a1 (s ) a1 (−s ) + r −1a2 (s ) a2 (−s )

(3.6.6)

Δ(s ) = det(sI − M )

(3.6.7)

where

and a1 (s ) and a2 (s ) are denominator and numerator polynomials of the open-loop transfer function, i.e., c(sI − A)−1 b =

a2 (s ) a1 (s )

(3.6.8)



139

It will be assumed that there are no common factors between a2 (s ) and a1 (s ) . Equation 3.6.6 establishes the fact that if s * is a root of Δ(s ) = 0, −s * would also be the root. Hence, eigenvalues of M are symmetric with respect to the imaginary axis of the s-plane. Recall that eigenvalues of any real matrix are always symmetric with respect to the real axis. Hence, eigenvalues of M are symmetric with respect to both real and imaginary axes. Furthermore, (−1) n Δ( jω ) = a1 ( jω ) a1 (− jω ) + r −1a2 ( jω ) a2 (− jω ) 2

= a1 ( jω ) + r −1 a2 ( jω )

(3.6.9)

2

Note that ai ( jω ) , i = 1 and 2, will be zero only when s = 0 is a root of ai (s ) = 0 . Because it has been assumed that there are no common factors between a1(s) and a2(s), |a1(jω)| and |a2(jω)| cannot simultaneously be zero. As a result, Δ( jω ) > 0 ; i.e., none of the eigenvalues is on the imaginary axis of the s-plane. Fact The optimal poles are the stable eigenvalues of M (Kailath, 1980).

THE SYMMETRIC ROOT LOCUS From (3.6.6), a1 (s ) a1 (−s ) + r −1a2 (s ) a2 (−s ) = 0

(3.6.10)

Hence,

r −1

a2 (s ) a2 (−s ) = −1 a1 (s ) a1 (−s )

(3.6.11)

Let m

a2 (s ) =

∏ (s − z ) i

(3.6.12)

i=1

n

a1 (s ) =

∏ (s − p ) i

i =1

(3.6.13)


140


m

∏ (s − z )(s + z ) i

−1

r (−1)

m−n

i

i =1 n

∏ (s − p )(s + p ) i

= −1

(3.6.14)

i

i=1

Equation 3.6.14 is in the standard root locus form (Kuo, 1995). When n − m is even, the 180° root locus plot will be constructed. If n − m is odd, the 0° root locus will be constructed. Case I: High Cost of Control (r → ∞) When the cost of control action is high, it is desired to use a small value of input u(t). This can be achieved by selecting a very large value of r. When r → ∞ , Equation 3.6.6 yields Δ(s ) = a1 (s ) a1 (−s ) = 0

(3.6.15)

Because the optimal poles are the stable roots of Δ(s ) = 0 , we have the following two situations: 1. Stable open-loop system: If all the eigenvalues of the matrix A are in the left half of the s-plane, optimal closed-loop poles are the same as openloop poles. 2. Unstable open-loop system: Optimal closed-loop poles are: (a) stable open-loop poles and (b) reflections of unstable open-loop poles about the imaginary axis. Case II: Low Cost of Control (r → 0) When r → 0 , Equation 3.6.10 yields a2 (s ) a2 (−s ) = 0

(3.6.16)

Hence, m optimal closed-loop poles are the stable roots of a2 (s ) a2 (−s ) = 0 . They are either open-loop left-half zeros or the reflections of open-loop right-half zeros about the imaginary axis. The remaining (n − m) optimal closed-loop poles are located near infinity. To find their locations, Equation 3.6.10 is written as follows for large s by ignoring lower powers of s: s n (−s ) n + r −1s m (−s )m (b0 )2 = 0 where

(3.6.17)



141

a2 ( s ) = b0 s m + … + bm

(3.6.18)

s 2 ( n −m ) = (−1) n −m +1 b02 r −1

(3.6.19)

From (3.6.17),

or 1

1

s = [(−1)( n −m +1) ]2 ( n −m ) (b02 r −1 ) 2 ( n −m )

(3.6.20)

n − m + 1 odd 1

1

s = (−1) 2 ( n −m ) (b02 r −1 ) 2 ( n −m )

(3.6.21)

Recall that −1 = e jπ(2 +1) , where is an integer Hence 1

j π ( 2 +1)

(−1) 2 ( n −m ) = e 2 ( n −m ) ;

= 0,1, 2,…, 2(n − m ) − 1

(3.6.22)

Substitution of (3.6.22) into (3.6.21) yields 2(n – m) roots of Equation 3.6.19. n − m + 1 even 1

1

s = (1) 2 ( n −m ) (b02 r −1 ) 2 ( n −m )

(3.6.23)

1 = e jπ 2 , where is an integer

(3.6.24)

Recall that

Hence 1

jπ 2

(1) 2 ( n −m ) = e 2 ( n −m ) ;

= 0, 1, 2,......., 2 ( n − m ) − 1

(3.6.25)


142


Substitution of (3.6.25) into (3.6.23) yields 2( n − m ) roots of Equation 3.6.19. To summarize, the 2(n – m) roots of equation (3.6.19) lie on a circle of radius 1

(b02 r −1 ) 2 ( n −m ) in a pattern described by (3.6.21) and (3.6.25). This pattern is known as Butterworth configuration (Kailath, 1980). EXAMPLE 3.12: NONCOLOCATED SENSOR

AND

ACTUATOR (BRYSON, 1979)

Consider the two-degree-of-freedom spring-mass system shown in Figure 3.11, where the force u0 (t ) is applied on the left mass and the position of the right mass y (t ) is the output. This model has been extensively used to simulate noncolocated sensor and actuator in the structural vibration control. The governing differential equations of motion are m0 x + k ( x − y ) = u0 (t )

(3.6.26)

m0 y + k ( y − x ) = 0

(3.6.27)

Nondimensional time t ′ is defined as k m0

t′ = t

(3.6.28)

Therefore, d (.) = dt ′

m0 d (.) k dt

and

d 2 (.) m0 d 2 (.) = k dt 2 dt ′ 2

(3.6.29a,b)

Hence, Equation 3.6.26 and Equation 3.6.27 can be written as x ′′ + ( x − y ) = u (t ′)

where u =

y ′′ + ( y − x ) = 0 x (t ) u0

k

FIGURE 3.11 A two-degree-of-freedom system.

(3.6.30)

(3.6.31) y (t )

m0

u0 k

m0



143

Taking the Laplace transform of (3.6.30) and (3.6.31) with zero initial conditions, s 2 x (s ) + ( x (s ) − y (s )) = u (s )

(3.6.32)

s 2 y (s ) + ( y (s ) − x (s )) = 0

(3.6.33)

Some simple algebra yields the SISO transfer function: y (s ) a (s ) 1 = = 2 u (s ) s 2 (s 2 + 2 ) a1 (s )

(3.6.34)

Hence, the root locus equation is r −1

a2 (s ) a2 (−s ) = −1 a1 (s ) a1 (−s )

(3.6.35)

r −1

1 = −1 s (s + 2 )2

(3.6.36)

or

4

2

The 180° root locus is shown in Figure 3.12. There are eight branches and all of them end at infinity. Angles of asymptotes are (2 l + 1)

π ; 8

l = 0, 1, 2, 3, 4, 5, 6, 7

(3.6.37)

All these asymptotes intersect the real axis at the origin of complex plane.

3.7 LINEAR QUADRATIC TRAJECTORY CONTROL It is desired to find an optimal control law in such a way as to cause the output y(t ) to track or follow a desired trajectory η(t ) . Hence, the objective function (3.5.2) is modified (Sage and White, 1977) to be tf

∫

1 1 I = z T (t f ) S f z (t f ) + [z T (t )Qz (t ) + u T (t ) Ru(t )]dt 2 2 0

where z(t ) is the trajectory error defined as

(3.7.1)


144


Symmetric Root Locus 3

2 x Imaginary Axis

1

0

–1 x –2

–3 –3

–2

–1

0 Real Axis

1

2

3

FIGURE 3.12 A symmetric root locus plot.

z(t ) = η(t ) − y(t )

(3.7.2)

The state space equation (3.5.1) is modified (Sage and White, 1977) to include a deterministic external input or the plant noise vector w (t ) : dx = Ax (t ) + Bu(t ) + w(t ); dt

x (0 ) = x 0

(3.7.3)

From Equation 3.4.11, the Hamiltonian H is defined as 1 H = [ z T (t )Qz(t ) + uT (t )Ru(t )] + λ T (t )[ Ax(t ) + Bu(t ) + w(t )] 2

(3.7.4)

The optimality condition (3.4.16) yields u(t ) = − R−1BT λ (t )

(3.7.5)

dλ = −{C T Q[Cx(t ) − η(t )] + AT λ (t )} dt

(3.7.6)

Equation 3.4.12 yields



145

with the terminal condition λ (t f ) = C T S (t f )[Cx(t f ) − η(t f )]

(3.7.7)

In order to determine the closed-loop control law, the transformation (3.5.14) is modified (Sage and White, 1977) to be λ (t ) = S (t )x(t ) − ξ(t )

(3.7.8)

where ξ(t ) is to be determined. Differentiating (3.7.8), d λ dS dx d ξ = x(t ) + S (t ) − dt dt dt dt

(3.7.9)

Using (3.7.3), (3.7.5), and (3.7.6), ⎛ dS −1 T T T ⎞ ⎜ + SA − SBR B S + C QC + A S ⎟ x(t ) ⎝ dt ⎠ ⎛ dξ ⎞ + ⎜− + SBR−1BT ξ + Sw(t ) − C T Qη(t ) − AT ξ(t )⎟ = 0 ⎝ dt ⎠

(3.7.10)

Using (3.7.8), the terminal condition (3.7.7) can be written as S (t f )x(t f ) − ξ(t f ) = C T S (t f )Cx(t f ) − C T S (t f )η(t f )

(3.7.11)

The solution of (3.7.10) can be obtained by solving it as two separate problems: dS + SA − SBR −1BT S + C T QC + AT S = 0 dt with S (t f ) = C T S (t f )C and − with

dξ + SBR−1BT ξ + Sw(t ) − C T Qη(t ) − AT ξ(t ) = 0 dt

(3.7.12)


146


ξ(t f ) = C T S (t f )η(t f )

(3.7.13)

Lastly, from (3.7.5) and (3.7.8), the control law is u(t ) = − R−1BT [ S (t )x(t ) − ξ(t )] = −K (t )x(t ) + R−1BT ξ(t )

(3.7.14)

The state feedback gain matrix K (t ) is the same as that given by (3.5.19). Hence, the solution of the linear quadratic trajectory control problem is composed of two parts: (1) a linear regulator part and (2) a correction term containing ξ(t ) . The computation of ξ(t ) requires the solution of (3.7.13) backward in time. Hence, it is required that w(t ) and η(t ) are exactly known a priori for all time t. From the disturbance rejection point of view, the control law can be described to be noncausal. EXAMPLE 3.13: OPTIMAL CONTROL (HUGHES, 1979)

OF

SUN TRACKING SOLAR CONCENTRATORS

A solar collector consists of a concentrator and a receiver. As a concentrator, point focusing parabolic dishes have been used. It reflects the sun’s energy towards its focal point and the receiver accepts the concentrated energy for further conversions. The axis of the paraboloid must be pointed at the sun in order to produce the required flux densities at the receiver aperture. Whenever there is a pointing error, energy is lost; hence, this energy loss is minimized by an appropriate control technique. A linear model of a single axis of the concentrator, which is driven by an electric motor, is shown in Figure 3.13, where θ0 (t ) : Collector’s line of sight (LOS) θi (t ) : Sun’s position u (t ) : Command input to the motor ξ: Damping ratio K s : System gain ωn: Natural frequency The state space model of the system can be written as

where

dx = Ax + bu (t ) dt

(3.7.15)

y (t ) = Cx (t )

(3.7.16)



147

i

u (t )

2 n

Ks s( s

2

2

ns

0

(t )

(t )

e (t )

2 n)

FIGURE 3.13 A model for sun tracking solar concentrator. 3

A

2

x2 (t)

1

0

O

–1

–2 B –3 –8

–6

–4

–2

0 x1 (t)

2

4

6

8

FIGURE 3.14 State space trajectories (___ u = 1, ---- u = 1).

⎡ θ0 ⎤ ⎢ ⎥ x = ⎢θ 0 ⎥ ; ⎥ ⎢θ ⎣ 0⎦ ⎡ 0 ⎤ ⎢ ⎥ b=⎢ 0 ⎥; ⎢ K s ω 2n ⎥ ⎣ ⎦

⎡0 ⎢ A = ⎢0 ⎢0 ⎣ and

1 0 −ω 2n ⎡1 ⎢ C = ⎢0 ⎢0 ⎣

0 ⎤ ⎥ 1 ⎥; −2 ξω n ⎥⎦ 0 1 0

0⎤ ⎥ 0⎥ 1 ⎥⎦

(3.7.17)

Let η(t ) be a vector of the sun’s position, velocity, and acceleration, i.e.,


148


⎡θi ⎤ ⎢ ⎥ η(t ) = ⎢θ i ⎥ ⎢ ⎥ ⎣θi ⎦

(3.7.18)

To minimize the energy loss, the output vector y(t ) should follow the sun’s trajectory η(t ) as closely as possible. Hence, a linear servomechanism or LQ tracking problem is solved. The objective function is

1 J= 2

t f − t0

∫ z Qz + ru dt T

2

(3.7.19)

0

where z(t ) = η(t ) − y(t )

(3.7.20)

t0 is the time of sunrise, and t f is the time of sunset. Using (3.7.14), the control law is given as u (t ) = −K (t )x(t ) + r −1bT ξ(t )

(3.7.21)

The variable ξ(t ) is obtained by solving Equation 3.7.13 with w(t ) = 0 . Because the desired trajectory η(t ) changes every day, the variable ξ(t ) has to be computed every day.

3.8 FREQUENCY-SHAPED LQ CONTROL Let the quadratic objective function be ∞

I=

∫ x(t) Qx(t) + u(t) Ru(t)dt T

T

(3.8.1)

0

Using Parseval’s Theorem (Appendix B),

I=

1 2π

∞

∫x

T

(− jω )Qx ( jω ) + u T (− jω ) Ru( jω ) d ω

(3.8.2)

−∞

Modify the objective function by making Q and R functions of the frequency ω (Gupta, 1980):



1 I= 2π

149

∞

∫x

T

(− jω )Q (ω ) x ( jω ) + u T (− jω ) R (ω )u( jω ) d ω

(3.8.3)

−∞

Assume that Q(ω ) and R(ω ) are rational functions of ω2 . Factoring these matrices, Q (ω ) = N qT (− jω ) N q ( jω )

(3.8.4)

R (ω ) = N rT (− jω ) N r ( jω )

(3.8.5)

Let the rank of the matrix Q(ω ) be p. Then, N q ( jω ) will be an p × n matrix. For the existence of the solution of an LQ problem, the matrix R(ω ) has to be of full rank. Therefore, N r ( jω ) will be an m × m matrix Define x T (− jω )Q (ω ) x ( jω ) = z T (− jω )z ( jω )

(3.8.6)

N q ( jω ) x ( jω ) = z ( jω )

(3.8.7)

uT (− jω)R(ω)u( jω) = χ T (− jω)χ( jω)

(3.8.8)

N r ( jω)u( jω) = χ( jω)

(3.8.9)

where

Similarly,

where

The objective function (3.8.3) can be expressed as

I=

1 2π

∞

∫ z (− jω)z( jω) + χ (− jω)χ( jω)dω T

T

(3.8.10)

−∞

Using Parseval’s Theorem (Appendix B), ∞

I=

∫ z (t )z(t ) + χ (t )χ(t )dt T

0

T

(3.8.11)


150


From (3.8.7), z(t ) is the output of the linear system with the transfer function matrix N q (s ) and input x(t ) ; i.e., z (s ) = N q (s ) x (s )

(3.8.12)

Similarly, from (3.8.9), χ(t ) is the output of the linear system with the transfer function matrix N r (s ) and input u(t ) ; i.e., χ( s ) = N r ( s )u( s )

(3.8.13)

Let the state space model of the MIMO system (3.8.12) be ξ = Aq ξ + Bq x(t )

(3.8.14)

z(t ) = Cq ξ + Dq x(t )

(3.8.15)

Similarly, let the state space model of the MIMO system (3.8.13) be ν = Ar ν + Br u(t )

(3.8.16)

χ(t ) = Cr ν + Dr u(t )

(3.8.17)

Define the augmented state vector x a (t ): ⎡x⎤ ⎢ ⎥ x a (t ) = ⎢ ξ ⎥ ⎢⎣ν⎥⎦

(3.8.18)

Let the augmented state space model be x a (t ) = Aa x a (t ) + Ba u(t )

(3.8.19)

0 ⎤ ⎥ 0 ⎥ A r ⎥⎦

(3.8.20)

where ⎡A ⎢ Aa = ⎢ Bq ⎢0 ⎣ Now, it can be shown that

0 Aq 0

and

⎡B⎤ ⎢ ⎥ Ba = ⎢ 0 ⎥ ⎢ Br ⎥ ⎣ ⎦



151

z T z + χ T χ = xTa Q f x a + 2 xTa N f u + uT R f u

(3.8.21)

where ⎡ DqT Dq ⎢ Q f = ⎢C qT Dq ⎢ ⎢⎣ 0

0 ⎤ ⎥ 0 ⎥, ⎥ C rT C r ⎦⎥

DqT C q T q

C Cq 0

⎡ 0 ⎤ ⎢ ⎥ Nf = ⎢ 0 ⎥ , ⎢C rT Dr ⎥ ⎣ ⎦

and

R f = DrT Dr (3.8.22)

+ 2 x Ta N f u + u T R f udt

(3.8.23)

Therefore, the objective function (3.8.11) is ∞

I=

∫x Q x T a

f

a

0

Using (3.5.29), the optimal control law will be u(t ) = − K a x a (t )

(3.8.24)

T K a = R −1 f ( PBa + N f )

(3.8.25)

where

and T P( Aa − Ba R−f 1 N Tf ) + ( Aa − Ba R−f 1 N Tf )T P − PBa R−f 1 BaT P + Q f − N f R−1 f N f = 0 (3.8.26)

Representing K a as K a = ⎡⎣ K x

Kξ

K ν ⎤⎦

u(t ) = −K x x(t ) − K ξ ξ(t ) − K ν ν(t )

(3.8.27) (3.8.28)

EXAMPLE 3.14: A SIMPLE HARMONIC OSCILLATOR Consider a second-order system: y + ω 2n y = u (t ) for which the state equations are

(3.8.29)


152


x = Ax + bu (t )

(3.8.30)

y (t ) = x1 (t )

(3.8.31)

where x = ⎡⎣ x1 ⎡ 0 A=⎢ 2 ⎣ −ω n

x2 ⎤⎦

1⎤ ⎥ 0⎦

T

and

(3.8.32) ⎡0 ⎤ b=⎢ ⎥ ⎣1⎦

(3.8.33a,b)

Standard LQ Control Let the objective function be ∞

I=

∫ (y

2

+ u 2 ) dt

(3.8.34)

0

The optimal state feedback control law is u (t ) = −kx (t )

(3.8.35)

where k is the state feedback gain vector. For ω n = 10 rad/sec, k = ⎡⎣0.005

0.1⎤⎦

(3.8.36)

and the eigenvalues of the optimal closed-loop system are −0.05 ± 10 j . Therefore, if the disturbance to the system is a sinusoidal function with the frequency equal to 10 rad/sec, its effects will be extremely large. Frequency-Shaped LQ Control Let the objective function be 1 I= 2π where

∞

∫ [ y ( jω )

−∞

2

2

2

Q ( jω ) + u ( jω ) ]d ω

(3.8.37)



153

ω 2n s s + 2 ςω n s + ω 2n

Q (s ) =

(3.8.38)

2

For a small value of the damping ratio ς , Q ( jω n ) will be extremely large, and as a result the response of the optimal closed-loop system is expected to be insensitive to external disturbance at the frequency ω = ω n . Now, let Q (s ) y (s ) = z(s )

(3.8.39)

The state space realization of the system (3.8.39) can be written as ⎡ ξ 1 ⎤ ⎡ 0 ⎢ ⎥ = ⎢ 2 ⎣ ξ 2 ⎦ ⎣ −ω n

1 ⎤ ⎡ ξ1 ⎤ ⎡ 0 ⎤ ⎥ ⎢ ⎥ + ⎢ ⎥ y (t ) −2 ςω n ⎦ ⎣ ξ 2 ⎦ ⎣ω 2n ⎦

(3.8.40)

and z(t ) = ξ 2 (t )

(3.8.41)

Combining (3.8.30) and (3.8.40), ⎡ x1 ⎤ ⎡ 0 ⎢ ⎥ ⎢ 2 ⎢ x2 ⎥ = ⎢ −ω n ⎢ ξ 1 ⎥ ⎢ 0 ⎢ ⎥ ⎢ 2 ⎢⎣ ξ 2 ⎥⎦ ⎢⎣ ω n

1 0 0 0

0 ⎤ ⎡ x1 ⎤ ⎡0 ⎤ ⎥⎢ ⎥ ⎢ ⎥ 0 ⎥ ⎢ x2 ⎥ ⎢ 1 ⎥ u (t ) + 1 ⎥ ⎢ ξ1 ⎥ ⎢ 0 ⎥ ⎥⎢ ⎥ ⎢ ⎥ −2 ςω n ⎥⎦ ⎢⎣ ξ 2 ⎥⎦ ⎢⎣0 ⎥⎦

0 0 0 −ω 2n

(3.8.42)

From (3.8.39), the objective function (3.8.37) can be written as ∞

I=

∫ (z

2

+ u 2 ) dt

(3.8.43)

0

Equation 3.8.42 and Equation 3.8.43 constitute a standard LQ control problem. The optimal state feedback control law is u (t ) = −k a [ x1

x2

ξ1

ξ 2 ]T

(3.8.44)

where k a is the optimal state feedback gain vector. For ω n = 10 rad/sec and ς = 0.1 , k a = [ 4.1717

2.8885

−4.1717

0.0178]

(3.8.45)


154


And the eigenvalues of the optimal closed-loop system are −1.217 ± 10.9791 j and −1.2272 ± 8.96921 j . Therefore, if the disturbance to the system is a sinusoidal function with frequency 10 rad/sec, its effect will be small.

3.9 MINIMUM-TIME CONTROL OF A LINEAR TIME-INVARIANT SYSTEM Consider the linear system dx = Ax(t ) + Bu(t ); x(0 ) = x 0 dt

(3.9.1)

It is desired to apply a control u(t ) , such that the system reaches the origin of the state space, in a minimum time when inputs must satisfy the following constraints: ui (t ) ≤ α i ;

i = 1, 2,…, m

(3.9.2)

If the final time is denoted by t f , the objective is to minimize the following objective function: tf

I = tf =

∫ 1dt

(3.9.3)

0

Therefore, the Hamiltonian H is H = 1 + λ T ( Ax + Bu)

(3.9.4)

dλ ∂H =− = − AT λ dt ∂x

(3.9.5)

The adjoint equations are

The solution of (3.9.5) is T

λ (t ) = e− A t λ (0 )

(3.9.6)

H = 1 + λ T (0 )e− At Ax + λ T (0 )e− At Bu

(3.9.7)




155

Therefore, ∂H = pi ∂ui

(3.9.8)

pi = λ T (0 )e− At b i

(3.9.9)

where

B = ⎡⎣b1

b2

.

.

b m ⎤⎦

(3.9.10)

u T = ⎡⎣u1

u2

.

.

um ⎤⎦

(3.9.11)

Using (3.4.17) and (3.4.18), the optimal control inputs are ⎧+α u0 i (t ) = ⎨ i ⎪⎩−α i

NORMALITY

OF

if if

pi < 0 pi > 0

(3.9.12)

LINEAR SYSTEMS

When pi (t ) = 0 over a finite interval [t1 t 2 ] , it is called a singular control case. This also implies that higher derivatives of pi (t ), with respect to time, are zero over this finite interval. Therefore, for t ∈ [t1, t 2 ] , pi = λ T (0 )e− At b i = 0 −

dpi = λ T (0 )e− At Ab i = 0 dt

d 2 pi = λ T (0 )e− At A 2 b i = 0 dt 2

(3.9.13)

(−1)n−1

d n−1 pi = λ T (0 )e− At A n−1b i = 0 dt n−1

Equation 3.9.13 can be expressed as λ T (0 )e− At Ci = 0 where

(3.9.14)


156


C i = [b i

Ab i

A2 b i

.

.

A n −1b i ]

(3.9.15)

is the controllability matrix with respect to ui . It should be noted that λ (0 ) ≠ 0 because H = 1 otherwise; i.e., H ≠ 0 and the solution will not be optimal (Section 3.4.2). Therefore, for Equation 3.9.14 to be true, rank (C i ) < n

(3.9.16)

This analysis implies that there are no finite intervals on which pi (t ) = 0, provided C i is not singular for any i, i = 1, 2, … m. A system for which C i is not singular for any i is called normal (Gopal, 1984).

EXISTENCE AND UNIQUENESS THEOREMS CONTROL (GOPAL, 1984)

ON

MINIMUM-TIME

1. If the linear time-invariant system is controllable, and if all the eigenvalues of A have nonpositive real parts, a time-optimal control exists that transfers any initial state to the origin of the state space in a minimum time. 2. If the linear time-invariant system is normal, and if the time-optimal control exists, it is unique. 3. If eigenvalues of the matrix A are real and a unique time-optimal exists, each control component can switch at the most (n–1) times, where n is the dimension of the state space. EXAMPLE 3.15: TIME-OPTIMAL CONTROL INTEGRATOR SYSTEM

OF A

RIGID BODY

OR A

DOUBLE

Consider the system dx = Ax (t ) + Bu (t ) dt

(3.9.17)

where ⎡0 A =⎢ ⎣0

1⎤ ⎥ 0⎦

and

⎡0 ⎤ B =⎢ ⎥ ⎣1 ⎦

(3.9.18)

and u ≤1

(3.9.19)



157

The objective is to find the optimal control u(t), such that the system reaches the origin of the state space in a minimum time from an initial state x(0 ) . First, note that eigenvalues of A are 0 and 0, which are real and nonpositive. Furthermore, ⎡0 C=⎢ ⎣1

1⎤ ⎥ 0⎦

(3.9.20)

Therefore, the system is controllable and normal as well. Application of the existence and uniqueness theorem indicates that a unique optimal control exists with at the most one switching. Hamiltonian H = 1 + λ1 x2 + λ 2 u

(3.9.21)

∂H = λ2 ∂u

(3.9.22)

Hence, the optimal control input is given by ⎧−1 u (t ) = ⎨ ⎩⎪+1

if if

λ2 > 0 λ2 < 0

(3.9.23)

Equation 3.9.23 can be expressed as u (t ) = − sgn(λ 2 (t ))

(3.9.24)

∂H d λ1 =− =0 dt ∂x1

(3.9.25)

d λ2 ∂H =− = − λ1 dt ∂x2

(3.9.26)

λ1 (t ) = λ1 (0 )

(3.9.27)

Adjoint equations are

The solution of (3.9.25) is



158


λ 2 (t ) = − λ1 (0 )t + λ 2 (0 )

(3.9.28)

Because states are specified at initial and final time, there are no constraints on adjoint variables. Hence, state equations must be used to determine adjoint variables. If u = 1 , the solution to (3.9.17) is x2 (t ) = t + x2 (0 )

x1 (t ) =

t2 + x2 (0 )t + x1 (0 ) 2

(3.9.29)

(3.9.30)

Eliminating t between (3.9.29) and (3.9.30),

x1 =

x22 ( x (0 ))2 + x1 (0 ) − 2 2 2

(3.9.31)

Equation 3.9.31 describes state space trajectories (which are parabolas) when u = 1 . These trajectories are shown as solid curves in Figure 3.14. If u = −1 , solution to (3.9.17) is x2 (t ) = −t + x2 (0 )

x1 (t ) = −

t2 + x2 (0 )t + x1 (0 ) 2

(3.9.32)

(3.9.33)

Eliminating t between (3.9.32) and (3.9.33),

x1 = −

x22 ( x (0 ))2 + x1 (0 ) + 2 2 2

(3.9.34)

Equation 3.9.34 describes state space trajectories (which are parabolas) when u = −1. These trajectories are shown as dashed curves in Figure 3.14. Because it is known that the optimal control input can only have one switching at the most, only four cases exist: Case I: u = 1 , 0 ≤ t ≤ tf

(3.9.35)



159

Initial states must be such that the system is on segment BO in Figure 3.14. Case II: u = −1 , 0 ≤ t ≤ t f

(3.9.36)

Initial states must be such that the system is on segment AO in Figure 3.14. Case III: ⎪⎧ +1 u=⎨ ⎪⎩−1

0 ≤ t ≤ ts ts < t ≤ t f

(3.9.37)

Initial states must be such that the system will follow one of solid parabolas, which intersect the segment AO. At the switching instant ts, the system will reach the segment AO and go to the origin of the state space along the segment AO. Case IV: ⎧⎪ −1 u=⎨ ⎩⎪+1

0 ≤ t ≤ ts ts < t ≤ t f

(3.9.38)

Initial states must be such that the system will follow one of dashed parabolas, which intersect the segment BO. At the switching instant ts , the system will reach the segment BO and go to the origin of the state space along the segment BO. The switching curve is composed of segments AO and BO (Figure 3.14). For segment AO,

x1 (0 ) +

( x2 (0 ))2 =0 2

(3.9.39)

Equation 3.9.39 implies the following equation for the curve AO: x1 (t ) = − For segment BO,

1 ( x2 (t ))2 2

(3.9.40)


160


x1 (0 ) −

( x2 (0 ))2 =0 2

(3.9.41)

Equation 3.9.41 implies the following equation for the curve BO: x1 (t ) =

1 ( x2 (t ))2 2

(3.9.42)

Combining (3.9.40) and (3.9.42), the equation of the switching curve AOB can be expressed as x1 (t ) = −

1 x2 (t ) x2 (t ) 2

(3.9.43)

On the basis of (3.9.43), a switching function (Gopal, 1984) is defined as follows: s ( x (t )) = x1 (t ) +

1 x2 (t ) x2 (t ) 2

(3.9.44)

If s > 0 , x(t) lies above the curve AOB. In this case, u = −1 because dotted parabolas are directed toward the switching curve. If s < 0 , x(t) lies below the curve AOB. In this case, u = 1 because solid parabolas are directed toward the switching curve. If s = 0 and x2 > 0 , x (t) lies on the curve AO. In this case, u = −1 . If s = 0 and x2 < 0 , x (t) lies on the curve BO. In this case, u = 1 . In summary, the time-optimal control in feedback form is ⎧−1 ⎪ ⎪+1 u (t ) = ⎨ ⎪+1 ⎪⎩−1

when when when when

s(x(t )) > 0 s(x(t )) = 0 s(x(t )) < 0 s(x(t )) = 0

and

x2 (t ) < 0

and

x2 (t ) > 0

(3.9.45)

The structure of the optimal feedback control system is shown in Figure 3.15. Control Law Eq. (3.9.45)

u(t )

x1

x2

x2

u( t )

x (t )

FIGURE 3.15 Implementation of bang-bang control via state feedback.



161

q1

q2

k u

m

m

b

FIGURE 3.16 A system with rigid and flexible modes.

EXAMPLE 3.16: MINIMUM-TIME CONTROL FLEXIBLE MODES

OF A

SYSTEM

WITH

RIGID

AND

The differential equations of motion of the mechanical system shown in Figure 3.16 are ⎡m ⎢ ⎣0

0 ⎤ ⎡ q1 ⎤ ⎡ b ⎥⎢ ⎥+ ⎢ m ⎦ ⎣q2 ⎦ ⎣ −b

−b ⎤ ⎡ q1 ⎤ ⎡ k ⎥⎢ ⎥+ ⎢ b ⎦ ⎣q2 ⎦ ⎣ − k

− k ⎤ ⎡ q1 ⎤ ⎡u (t ) ⎤ ⎥⎢ ⎥ = ⎢ ⎥ k ⎦ ⎣q2 ⎦ ⎣ 0 ⎦

(3.9.46)

Natural frequencies are found to be ω 20 = 0

(3.9.47)

2k m

(3.9.48)

and

ω 12 = Modal vectors are as follows: ⎡⎣1

1⎤⎦

T

corresponding to ω 20 = 0 (Rigid mode)

(3.9.49)

corresponding to ω 12 = 2 k / m (Flexible mode)

(3.9.50)

and ⎡⎣1

−1⎤⎦

T

Now, the displacement vector can be represented as a linear combination of these modal vectors: ⎡ q1 ⎤ ⎡ r1 ⎤ ⎢ ⎥ = Φ⎢ ⎥ ⎣q2 ⎦ ⎣r2 ⎦

(3.9.51)


162


where ⎡1 Φ=⎢ ⎣1

1⎤ ⎥ −1⎦

(3.9.52)

Substituting (3.9.51) into (3.9.46), and premultiplying both sides by ΦT ,

r1 =

1 u 2m

(3.9.53)

and

r2 + 2 ςω1r2 + ω12 r2 =

1 u 2m

(3.9.54)

where ς = 2 b / (2 m ω1 ) . Defining state variables as x1 = r1 , x2 = r1 , x3 = r2 , and x 4 = r2

(3.9.55)

x = Ax (t ) + bu (t )

(3.9.56)

State equations are

where ⎡0 ⎢ 0 A=⎢ ⎢0 ⎢ ⎢⎣0

1 0 0 0

0 0 0 −ω 12

0 ⎤ ⎥ 0 ⎥ 1 ⎥ ⎥ −2 ςω 1 ⎥⎦

and

⎡ 0 ⎤ ⎥ ⎢ 0.5m ⎥ b=⎢ ⎢ 0 ⎥ ⎥ ⎢ ⎢⎣0.5m ⎥⎦

and

x(t f ) = 0

(3.9.57)

Let the initial and final states be x (0 ) = ⎡⎣ − L and

0

0

0 ⎤⎦

T

(3.9.58a and b)



163

u (t ) ≤ α

(3.9.59)

The minimum time control will be bang-bang and unique. However, because all eigenvalues are not real, the maximum number of switching is not necessarily (n–1). Let switching instants be ti , i = 1, 2, …, k, and initially u (t ) = α as the position of the system has to be increased. Solving state equations with this input and initial condition (3.9.58a), and then imposing the final condition (3.9.58b), the following nonlinear equations (Pao and Singhose,1998; Singhose and Pao, 1997) are obtained: k

(−1) k t f + 2

∑ (−1)

i−1

ti

(3.9.60)

i=1

k

(−1)

k +1 2 f

t +2

∑ (−1) (t ) i

=

2

i

i =1

4mL α

(3.9.61)

k

1 + (−1) k +1 e

ςω1t f

cos(ω d t f ) + 2

∑ (−1) e

i ςω1ti

cos(ω d ti ) = 0

(3.9.62)

i =1

k

(−1) k +1 e

ςω1t f

sin(ω d t f ) + 2

∑ (−1) e

i ςω1ti

sin(ω d ti ) = 0

(3.9.63)

i =1

where ω d = ω n 1 − ς 2 . There are many solutions of ti , i = 1, 2, …, k, that satisfy (3.9.60) to (3.9.63). However, only one of them will satisfy Equation 3.9.12. EXAMPLE 3.17: INPUT SHAPING Consider the prototype second-order system (Kuo, 1995): y + 2 ξω n y + ω 2n y = ω 2n u (t )

(3.9.64)

where ξ , ω n , and u(t) are the damping ratio, undamped natural frequency, and the reference input, respectively. The reference input is often a unit step function for which the response of an underdamped system is y (t ) = 1 − e − ξω nt cos ω d t − χe − ξω nt sin ω d t ;

t≥0

(3.9.65)


164


u(t)

1 Ap Ap 1

A2 A1 t1

t2

tp 1

tp

t

FIGURE 3.17 A staircase input.

where χ=

ξ 1 − ξ2

and

ω d = ω n 1 − ξ2

(3.9.66a,b)

The value of the steady state error is guaranteed to be zero. However, there is a large amount of oscillation, and the settling time can be large for a small damping level. The concept of input shaping is to modify the command input so that that the system reaches the final position with zero oscillation. For the staircase input shown in Figure 3.17, the response can be written as p

y (t ) =

∑ A [1 − e i

− ξω n ( t − ti )

(cos ω d (t − ti ) + χ sin ω d (t − ti ))]us (t − ti ) (3.9.67)

i =1

where us (t − ti ) is a unit step function applied at t = ti . For t > t p , p

y (t ) = g (t ) +

∑A

i

i =1

where

(3.9.68)



165

g (t ) = −e −ξω nt [α sin ω d t + β cos ω d t ]

(3.9.69)

p

α=

∑e

ξω n ti

Ai (sin ω d ti + χ cos ω d ti )

(3.9.70)

ξω n ti

Ai (cos ω d ti − χ sin ω d ti )

(3.9.71)

i=1

p

β=

∑e i=1

The conditions for g(t) to be zero for t > t p are p

α=

∑e

ξω n ti

Ai (sin ω d ti + χ cos ω d ti ) = 0

(3.9.72)

ξω n ti

Ai (cos ω d ti − χ sin ω d ti ) = 0

(3.9.73)

i=1

p

β=

∑e i=1

Equation 3.9.72 and Equation 3.9.73 are equivalent to the following well-known conditions (Singer and Seering, 1990): p

∑e

ξω n ti

Ai sin ω d ti = 0

(3.9.74)

ξω n ti

Ai cos ω d ti = 0

(3.9.75)

i=1

p

∑e i=1

Consider the case of p = 2 . Without any loss of generality, assume that t1 = 0

and

A1 + A2 = 1

(3.9.76a and b)

In this case, Equation 3.9.74 and Equation 3.9.75 can be solved for two unknowns. The results are


166


1 π t2 = , A1 = , and K = e 1+ K ωd

−

ξπ 1− ξ2

(3.9.77a, b, and c)

Therefore, from (3.9.68), y (t ) = 1

for t > t2

(3.9.78)

In other words, the system reaches the desired value without any oscillation for t > t2 , when the input u(t) is as follows: ⎧⎪(1 + K )−1; 0 < t ≤ π / ω d u (t ) = ⎨ 1; t > π / ω d ⎩⎪

(3.9.79)

Time Optimality of the Shaped Input Define x2 = y ,

x1 = y − 1 ,

v = u −1

(3.9.80)

The state equations are x = Ax + bv (t )

(3.9.81)

where x = ⎡⎣ x1 ⎡ 0 A=⎢ 2 ⎣ −ω n

1 ⎤ ⎥ −2 ξω n ⎦

x2 ⎤⎦

T

and

(3.9.82) ⎡0 ⎤ b=⎢ 2⎥ ⎣ω n ⎦

(3.9.83a,b)

Now, initial and final states are x(0 ) = ⎡⎣ −1

0 ⎤⎦

T

and

x(t f ) = ⎡⎣0

0 ⎤⎦

T

(3.9.84a,b)

where t f is the final time. Another constraint for a positive shaper is that the input u(t) should be between 0 and 1. This implies the following constraint on v(t): −1 ≤ v (t ) ≤ 0

(3.9.85)



167

For minimum-time control, the Hamiltonian H is defined as follows: H = 1 + λ1 x2 + λ 2 (−ω 2n x1 − 2 ξω n x2 + ω 2n v )

(3.9.86)

where adjoint variables λ1 (t ) and λ 2 (t ) satisfy the following equations: ∂H λ 1 = − = ω 2n λ 2 ∂x1

(3.9.87)

∂H = − λ1 + 2 ξω n λ 2 λ 2 = − ∂x2

(3.9.88)

∂H = ω 2n λ 2 ∂v

(3.9.89)

and

Therefore, the minimum-time control will depend on the sign of λ 2 (t ). Differentiating (3.9.88) and using (3.9.87), − 2 ξω λ + ω 2 λ = 0 λ 2 n 2 n 2

(3.9.90)

⎤ ⎡ λ (0 ) − ξωn λ 2 (0 ) λ 2 (t ) = eξωnt ⎢ 2 sin ωd t + λ 2 (0 )cos ωd t ⎥ ωd ⎦ ⎣

(3.9.91)

λ 2 (0 ) = − λ1 (0 ) + 2 ξω n λ 2 (0 )

(3.9.92)

Solving (3.9.90),

From (3.9.88),

Substituting (3.9.92) into (3.9.91), ⎡ −λ (0 ) + ξωn λ 2 (0 ) ⎤ λ 2 (t ) = eξωnt ⎢ 1 sin ωd t + λ 2 (0 )cos ωd t ⎥ ωd ⎣ ⎦

(3.9.93)

Because both x(0) and x(t f ) are specified, λ1 (0 ) and λ 2 (0 ) are not specified, and have to be determined so that the resulting input leads to satisfaction of initial


168


and final conditions of states. To achieve the control input (3.9.79) that leads to satisfaction of initial and final conditions of states, λ 2 (0 ) = 0

and

λ1 (0 ) < 0

(3.9.94)

Then, from (3.9.93), λ 2 (t ) = − λ1 (0 )e ξω nt sin ω d t > 0

for 0 < t
0

0

Construct the symmetric root locus plot to determine optimal closed-loop pole locations. Show the performance of your controller as r changes from 0 to ∞. Parameters (ECP Manual): J1 = 0.0024 kg-m2, J2 = 0.0019 kg-m2, J3 = 0.0019 kg-m2 k1 = k2 = 2.8 N-m/rad c1 = 0.007 N-m/rad/sec, c2 = c3 = 0.001 N-m/rad/sec khw =

100 units 6

P3.19 Refer to Example 3.11. Select q1 = 10 and q2 = 1. a. Find optimal state feedback gain for three different values of ρ: 10 −7 , 10 −8 , and 10 −9 . In each case, assume that the initial condition on state vector is −[1 1 0 0 ]T and plots x1 (t ) , x2 (t ) , and u(t) vs. time. Discuss your results. b. Refer to Equation 3.5.88 and Figure 3.10. For all the three aforementioned values of ρ, find λ 2 and β2 . Also, plot f(t) vs. time for initial state vector chosen in part a. Discuss your results. P3.20 The following model is developed for longitudinal pressure oscillation in a uniform chamber (Yang et al., 1992): n

i + ω i2 ηi + η

∑ (D η + E η ) = v (t) ; i = 1, 2, …, n i

=1

i

i



2a 2 vi (t ) = α

177

m

∑ ψ (z i

=1

a

)

u (t ) p

where p and u (t ) are the mean pressure and the pressure excitation supplied by the actuator# located at za , respectively. The normal mode# i of the chamber is described by ψi ( z ) = cos(i πz / α); 0 ≤ z ≤ α where α is the length of the chamber. Di i i i i

= = = =

1 2 3 4

Ei i i i i

= = = =

1 2 3 4

=1

=2

=3

=4

0.007 0.1 0.01 –0.005

–0.001 0.007 0.75 0.01

0.007 –0.001 0.008 1.50

=1

=2

=3

=4

–0.005 –0.0025 –0.005 0.01

–0.005 –0.015 0.0 0.02

0.0025 0.01 –0.02 0.02

0.0016 0.01 0.02 –0.025

–0.01 0.01 –0.01 0.02

a. Develop a state space realization and find the open-loop poles with n = 4. b. Select a suitable location of a single actuator with n = 4, and find the state feedback gain vector to locate the poles such that closed-loop frequencies are same as those of the open loop, and there is at least a 5% damping ratio in each mode. c. With the location of the actuator in part b, draw the symmetric root locus for the following objective function: ∞

⎛ ⎜ ⎜ ⎝

n

∫∑ 0

i =1

⎞ ηi ⎟ + ru12 dt ⎟ ⎠ 2

d. Select suitable locations for two actuators, and find a state feedback gain matrix to locate the poles, such that closed-loop frequencies are same as those of open loop, and there is at least a 5% damping ratio in each mode. e. With the location of the actuator in part b, draw the symmetric root locus for the following objective function:


178


∞

⎡⎛ ⎢⎜ ⎢⎜ ⎢⎣⎝

n

∫ ∑ 0

i=1

2 ⎤ ⎞ 2 2⎥ ⎟ ηi + u1 + u2 ⎥dt ⎟ ⎠ ⎥⎦

P3.21 Consider the flexible tetrahedral truss structure (Appendix F). The controller is to be designed on the basis of first four modes of vibration. Design a full state feedback LQR controller such that 1 2

∞

∫y

T

(t )y (t ) + ρu T (t )u(t ) dt

0

is minimized. Demonstrate your controller performance for two values of ρ = 0.1 and 1. P3.22 Consider the single-link flexible manipulator described in Problem P3.4. Consider two vibratory modes for the link. It is given that u < 1 N-m. If the link is to be rotated by 30°, find the switching instants and move time for the bang-bang (minimum-time) control. The link starts from the rest and there should not be any vibration when the arm reaches its final position.


4

Control with Estimated States

In Chapter 3, a state feedback control system was developed under the assumption that all states are available to us. In general, it will be impractical to measure all the states. Therefore, methods to estimate states are developed on the basis of inputs and outputs of a deterministic and a stochastic system as well. Then, theories and examples of optimal state estimation and linear quadratic Gaussian control are presented.

4.1 OPEN-LOOP OBSERVER Consider the system x = Ax (t ) + Bu(t )

(4.1.1)

If the initial conditions are estimated to be xˆ (0 ) = x(0 ) + ε

(4.1.2)

where ε is the error in the estimate of the initial state vector. Then, states can be computed by the solution of the following equation (Kailath, 1980): xˆ = Axˆ + Bu(t )

(4.1.3)

x = x (t ) − xˆ (t )

(4.1.4)

Define

Errors in the state vector can be obtained as the solution of x = Ax (t );

x (0 ) = −εε

(4.1.5)

Hence, if the matrix A is stable, x (t ) → 0 as t → ∞ . On the other hand, if the matrix A is unstable, x (t ) → ∞ as t → ∞ .

179


180


This strategy of estimating states will be useful only when all the eigenvalues of the matrix are far inside the left half of the s-plane because only in this case will errors die out quickly. But, in most cases that need control, the eigenvalues of the matrix A will be close to the imaginary axis, or the matrix A will be unstable. In these cases, initial state errors will either persist for a long time or lead to unbounded state errors.

4.2 CLOSED-LOOP OBSERVER The algorithm presented in Section 4.1 works in an open-loop fashion. Measured outputs, y(t ), are not used at all. At any instant, estimated outputs are yˆ (t ) = Cxˆ (t )

(4.2.1)

δy (t ) = y (t ) − yˆ (t )

(4.2.2)

Define an error signal,

If there is no estimation error, δy(t ) = 0 . The signal δy(t ) can be used as feedback variables to influence the dynamics of estimated states as follows (Kailath, 1980): xˆ = Axˆ + Bu(t ) + L ( y (t ) − yˆ (t )) ;

xˆ (0 ) = xˆ 0

(4.2.3)

where L is an observer gain matrix (Figure 4.1). Here, the following questions arise: 1. What is the role of L in having estimated states xˆ (t ) converge to true states x(t ) ? 2. Is it possible to select L for a guaranteed convergence of estimated states xˆ (t ) to true states x(t )? 3. Can we select L to achieve a desired rate of convergence of estimated states xˆ (t ) to true states x(t ) ? 4. What should be the value of xˆ (0 )? To answer these questions, the estimator equation (4.2.3) is subtracted from the state equation (4.1.1): x = Ax − L ( y (t ) − Cxˆ (t )) = ( A − LC ) x ;

x (0 ) = x (0 ) − xˆ (0 )

(4.2.4)

If the matrix ( A − LC ) is stable, lim x (t ) → 0 as t → ∞ , or x (t ) → xˆ (t ) as t → ∞ . Equation 4.2.3 is also called the Luenberger observer. Usually, xˆ (0 ) = 0 .



181

u(t )

y (t )

plant xˆ (t )

yˆ (t )

xà (t )

B

C

A

y (t )

y (t )

yˆ (t )

L

FIGURE 4.1 Schematic drawing of the closed-loop observer.

Theorem Eigenvalues of ( A − LC ) can be arbitrarily located in the complex plane, provided the realization ( A, C ) is observable. Proof As eigenvalues of a square matrix and its transpose are identical, det(sI − ( A − LC )) = det(sI − ( A − LC )T ) = det(sI − ( AT − C T LT ))

(4.2.5)

With the interpretation (Kailath, 1980) A → AT ;

B → CT ;

and

K → LT

(4.2.6)

the problem of finding L is dual to that of finding K (see Chapter 3). Therefore, eigenvalues of ( AT − C T LT ) can be arbitrarily placed in the complex plane, provided ( AT , C T ) is controllable. It is easy to see that the controllability of ( AT , C T ) is equivalent to the observability of ( A, C ) . This completes the proof. Definition 4.1.1: Detectability The detectability of a system is a weaker condition than the observability. It only requires that all the unstable modes of the system are observable. The system (A, C) is called detectable if there exists a matrix L such that the matrix (A-LC) is stable (Zhou et al., 1996).

4.2.1 DETERMINATION OF OBSERVER GAIN VECTOR L SINGLE-OUTPUT SYSTEM

FOR A

Let α be the vector of the coefficients of desired characteristic polynomial of the matrix ( A − Lc) ; i.e.,


182


α = [α1

α2

.

.

αn ]

(4.2.7)

det( sI − ( A − Lc)) = s n + α1s n−1 + … + α n−1s + α n

(4.2.8)

(α − a ) = LT C*Ut

(4.2.9)

Therefore,

where C* = [cT

AT cT

.

.

( AT ) n −1 cT ] = O T

(4.2.10)

where O is the observability matrix. Hence, from Equation 3.1.18, LT = (α − a )(OT Ut )−1 = (α − a )(Ut )−1 (OT )−1

(4.2.11)

Because the matrix U t is nonsingular, the solution of L can be obtained, provided O T or, equivalently, O is nonsingular. In other words, the existence of L is guaranteed for an arbitrary choice of α , provided the realization ( A, C ) is observable.

4.2.2 DETERMINATION OF OBSERVER GAIN MATRIX L MULTIPLE-OUTPUT SYSTEM

FOR A

For a multioutput system, the gain matrix L is not unique for specified locations of observer poles. A method, dual to that presented in Chapter 3, Section 3.2, can be developed to determine the gain matrix L.

4.2.3 LOCATIONS

OF

OBSERVER POLES

When observer poles are located farther into the left half of the s-plane, errors in estimated states go to zero at a faster rate. In general, this will also imply higher values of observer gains, which do not cost anything in terms of the actuator size or the magnitude of the control input. However, the observer starts behaving like a differentiator as the poles are pushed farther into the left half of the s-plane. As a result, high-frequency stochastic noises in the system can get amplified, and the system performance can be reduced. Therefore, the optimal selection of observer poles requires considerations of stochastic disturbances in the system, and will be presented in Section 4.6 as the Kalman filter (Kalman and Bucy, 1961). EXAMPLE 4.1: AN AORTIC PRESSURE OBSERVER

FOR AN

ARTIFICIAL HEART

The Penn State electric ventricular assist device (EVAD) consists of a high torque brushless DC motor coupled to a pusher plate by a roller-screw mechanism (Tsach



183

Blood Flow

Brushless DC Motor roller

Flexible Blood Sac

screw

Pusher Plate

FIGURE 4.2 Schematic drawing for Penn State electric ventricular assist device (EVAD).

et al., 1990). The blood flow is generated by moving the pusher plate against a flexible blood sac (Figure 4.2). The state space equations can be written as x = Ax (t ) + bu (t ) ⎡ x1 (t ) ⎤ ⎡0 ⎢ ⎥ ⎢ x(t ) = ⎢ x2 (t ) ⎥ A = ⎢0 ⎢ x3 (t ) ⎥ ⎢0 ⎣ ⎦ ⎣

1 −68.3 3.2

(4.2.12)

⎡ 0 ⎤ 0 ⎤ ⎥ ⎢ ⎥ −7.2 ⎥ b = ⎢ 425.4 ⎥ (4.2.13a,b,c) ⎢ 0 ⎥ −0.7 ⎥⎦ ⎣ ⎦

where states x1, x2 , and x3 are pusher plate’s position (mm), velocity (mm/sec), and aortic pressure (mm Hg), respectively. The input, u (t ) , is the DC motor voltage. The position of the pusher plate is measured by a sensor; i.e., the output is x1, and the output equation is y (t ) = cx (t )

(4.2.14)

where c = [1

0

0]

(4.2.15)

In this case, eigenvalues of A are 0, –1.0426, and –67.9574. The open-loop characteristic polynomial is det(sI − A) = s 3 + 69s 2 + 70.85s The observability matrix is

(4.2.16)


184


⎡1 ⎢ O = ⎢0 ⎢0 ⎣

0 1 −68.3

0 ⎤ ⎥ 0 ⎥ −7.2 ⎥⎦

(4.2.17)

As det O ≠ 0 , the state space realization is observable. Now, det( sI − A + Lc) = s 3 + (69 + 1 )s 2 + ( 70.85 + 69 1 + 2 )s + 70.85 1 + 0.7 2 − 7.2 3

(4.2.18)

where L = ⎡⎣ 1

2

3 ⎤⎦

T

(4.2.19)

Let the eigenvalues of the observer be –3, –27, and –67.9574. Then, the desired characteristic polynomial is det(sI − A + Lc) = s 3 + 98s 2 + 2119.7s + 5504.5

(4.2.20)

Comparing coefficients, 69 + 1 = 98

(4.2.21)

70.85 + 69 1 + 2 = 2119.7

(4.2.22)

70.85 1 + 0.7 2 − 7.2 3 = 5504.5

(4.2.23)

1 = 29 , 2 = 47.5 , and 3 = −474.527

(4.2.24)

Therefore,

EXAMPLE 4.2: A DISTURBANCE OBSERVER Consider a spring-mass-damper system (Chapter 2, Figure 2.4) subjected to a sinusoidal excitation. Then, the state equation (2.5.16) will be modified to be x = Ax (t ) + bu (t ) + bw (t ) where

(4.2.25)



185

w (t ) = w0 sin(ωt + φ)

(4.2.26)

It is assumed that the excitation frequency ω is known. However, the amplitude w0 and phase φ are not known. The objective is to design an observer to determine w(t) or, equivalently, the amplitude w0 and the phase φ . The excitation w(t) satisfies the following second-order equation: + ω 2 w = 0 w

(4.2.27)

Converting (4.2.27) into a state space model, ⎡ w 1 ⎤ ⎡ 0 ⎢ ⎥=⎢ 2 ⎣w 2 ⎦ ⎣ −ω

1 ⎤ ⎡ w1 ⎤ ⎥⎢ ⎥ ; 0 ⎦ ⎣w2 ⎦

w1 = w

(4.2.28)

Combining (4.2.25) and (4.2.28), x a = Aa x a (t ) + b a u (t )

(4.2.29)

where ⎡ x1 ⎤ ⎢ ⎥ x xa = ⎢ 2 ⎥ ; ⎢ w1 ⎥ ⎢ ⎥ ⎢⎣w2 ⎥⎦

⎡ 0 ⎢ −β / m Aa = ⎢ ⎢ 0 ⎢ ⎢⎣ 0

1

0 1 0

−α / m 0 0

−ω 2

0⎤ ⎥ 0⎥ ; 1⎥ ⎥ 0 ⎥⎦

⎡0 ⎤ ⎢ ⎥ 1 ba = ⎢ ⎥ ⎢0 ⎥ ⎢ ⎥ ⎢⎣0 ⎥⎦

(4.2.30a,b,c)

If the position of the mass is measured by a sensor, the output equation is y (t ) = cx a (t )

(4.2.31)

where c = [1

0

0

(4.2.32)

0]

The observability matrix is ⎡ 1 ⎢ 0 O=⎢ ⎢ −β / m ⎢ ⎢⎣αβ / m 2

0 1 −α / m −(β / m ) + (α 2 / m 2 )

0 0 1 −α / m

0⎤ ⎥ 0⎥ 0⎥ ⎥ 1 ⎥⎦

(4.2.33)


186


Therefore, det O = 1

(4.2.34)

And the realization is observable for all system parameters. Let m = 1 kg, α = 1 Nsec/m, β = 1N /m , and ω = 1 rad/sec. In this case, eigenvalues of Aa are –0.5 ± 0.86603j and ±1j . The open-loop characteristic polynomial is det(sI − Aa ) = s 4 + s 3 + 2 s 2 + s + 1 a = [1

1

1

1]

(4.2.35) (4.2.36)

Let the eigenvalues of the observer be −0.5 ± 0.86603 j and −0.5 ± 1 j . Then, the desired characteristic polynomial is det(sI − Aa + Lc) = s 4 + 2 s 3 + 3.25s 2 + 2.25s + 1.25

(4.2.37)

and α = [2

3.25

2.25

1.25 ]

(4.2.38)

Using Equation 4.2.11, LT = [1

0.25

0.25

−1]

Simulation results are shown in Figure 4.3, in which the estimated disturbance is shown by the dotted line.

4.3 COMBINED OBSERVER–CONTROLLER In Chapter 3, we designed the control system with the full state feedback. Because all states are not available, we can only feed back estimated states. Let the control law be defined (Figure 4.4) as u (t ) = v (t ) − K xˆ (t )

(4.3.1)

xˆ = Axˆ + Bu(t ) + L ( y (t ) − yˆ (t ))

(4.3.2)

where



1

187

w(t) we(t)

0.8

Disturbance and its Estimate

0.6 0.4 0.2 0 –0.2 –0.4 –0.6 –0.8 –1 0

5

10

15

20

25

Time (sec)

FIGURE 4.3 Estimation of the disturbance. u(t )

v (t )

y (t ) plant

B xˆ (t )

xˆ

K

L

A yˆ (t )

C

FIGURE 4.4 Combined observer–controller system.

Recall the plant dynamics x = Ax + Bu(t )

(4.3.3)

Subtracting (4.3.2) from (4.3.3), the state error x = x − xˆ is still governed by


188


x = ( A − LC ) x

(4.3.4)

x = ( A − BK ) x + BK x + Bv (t )

(4.3.5)

From (4.3.1) and (4.3.3),

Representing Equation 4.3.4 and Equation 4.3.5 in the matrix form, ⎡ x ⎤ ⎡( A − BK ) ⎢⎥ = ⎢ 0 ⎣ x ⎦ ⎣

BK ⎤ ⎡ x ⎤ ⎡ B ⎤ ⎥ ⎢ ⎥ + ⎢ ⎥ v(t ) ( A − LC ) ⎦ ⎣ x ⎦ ⎣ 0 ⎦

(4.3.6)

Equation 4.3.6 represents the dynamics of the closed-loop system (Figure 4.4). Note that the order of the closed-loop system is 2n where the dynamics of the plant is of the order n and observer dynamics is also of the order n. The system matrix for the closed-loop system is ⎡( A − BK ) Asys = ⎢ 0 ⎣

BK ⎤ ⎥ A − LC ⎦

(4.3.7)

Therefore, ⎡sI − ( A − BK ) det(sI 2 n − Asys ) = det ⎢ n 0 ⎣

⎤ − BK ⎥ sI n − ( A − LC ) ⎦

(4.3.8)

= det(sI n − ( A − BK )) det(sI n − ( A − LC )) The eigenvalues of the closed-loop system, Asys, are same as those of ( A − BK ) and ( A − LC ). Hence, K and L are to be chosen such that eigenvalues of ( A − BK ) and ( A − LC ) are located at desired places in the left half of the complex plane, respectively. In other words, K can be chosen as if all states are available for feedback, and L can be chosen as if only an observer as described in Section 4.2 is to be designed. This is called the separation property, which implies that the observer and controller can be designed independent of each other and the stability of the closed-loop system is guaranteed. However, as far as the performance of the closedloop system is concerned, there is certainly a coupling between selection of K and L. A rule of thumb that has often been used is to have eigenvalues of ( A − LC ) at least four times faster than those of ( A − BK ). But, more precisely, they can be selected to minimize certain performance criteria; e.g., LQG, H ∞ control, etc. The control law, (4.3.1) and (4.3.2), can be represented as a classical output feedback control system shown in Figure 4.5. Assuming v(t ) = 0 , Equation 4.3.1 and Equation 4.3.2 can also be written as



v(t )

189

y (t )

0

G (s )

u (t ) H (s )

FIGURE 4.5 Combined observer–controller system as an output feedback compensator.

xˆ = ( A − BK − LC ) xˆ + Ly (t )

(4.3.9)

u (t ) = − K xˆ (t )

(4.3.10)

and

Hence, the controller transfer function H (s ) can be expressed as H (s ) = − K (sI − ( A − BK − LC ))−1 L

(4.3.11)

It is important to realize that the stability of H (s ) or, equivalently, ( A − BK − LC ) is not guaranteed. It is possible to have an unstable ( A − BK − LC ) while ( A − BK ) and ( A − LC ) are stable. The transfer function of the closed-loop system is

y (s ) = ⎡⎣C

⎡B⎤ 0 ⎤⎦ (sI 2 n − Asys )−1 ⎢ ⎥ v (s ) ⎣0 ⎦

(4.3.12)

Because of the block-diagonal structure of Asys, ⎡(sI − ( A − BK ))−1 (sI 2 n − Asys )−1 = ⎢ n 0 ⎣

⎤ ? −1 1⎥ (sI n − ( A − Lc )) ⎦

(4.3.13)

Substituting (4.3.13) into (4.3.12), y (s ) = C (sI n − ( A − BK ))−1 Bv (s )

(4.3.14)

This transfer function does not depend on the observer dynamics (Kailath, 1980), and it is exactly the same as that obtained with the full state feedback in Chapter 3.


190


This can be understood by realizing that a transfer function is defined for zero initial conditions. In this case, x (0 ) = 0 , and the solution of Equation 4.3.4 yields x (t ) = 0

for all t ≥ 0

(4.3.15)

Therefore, an observer is not needed at all to estimate states. This result also highlights what an observer really guarantees. It basically drives the state errors caused by errors in initial values of states to zero. It may not compensate for modeling errors and unknown external disturbances. EXAMPLE 4.3: BALANCING

A

POINTER

Consider Example 3.1, Balancing a Pointer, in Chapter 3. If the length of the pointer is 0.98 m, ⎡0 A=⎢ ⎣10

1⎤ ⎥ 0⎦

and

⎡0⎤ b=⎢ ⎥ ⎣ −1⎦

(4.3.16)

Let the output y (t ) be the angular position of the pointer φ(t ); i.e., c = ⎡⎣1

0 ⎤⎦

(4.3.17)

To locate the controller poles at –1 and –1, Equation 3.1.32 yields k = ⎡⎣ −11

−2 ⎤⎦

(4.3.18)

Next, det(sI − ( A − Lc )) = s 2 + 1s + ( 2 − 10 )

(4.3.19)

Let the observer poles be at –4 and –4. Then, the desired characteristic equation will be det(sI − ( A − Lc )) = (s + 4)2 = s 2 + 8s + 16

(4.3.20)

Comparing coefficients of polynomials (4.3.19) and (4.3.20), 1 = 8

and

2 = 16

(4.3.21)

The controller equations (4.3.9 and 4.3.10), which are to be solved in real time, are as follows:



191

4 xˆ1

11

xˆ1

4 y (t )

u(t )

17 16 xˆ 2

2

xˆ 2

2

FIGURE 4.6 Observer-based state feedback controller to balance a pointer.

u (t ) = 11xˆ1 (t ) + 2 xˆ2 (t )

(4.3.22)

where xˆ1 (t ) = −4 xˆ1 + xˆ2 (t ) + 4 y (t ) ; xˆ2 (t ) = −17 xˆ1 − 2 xˆ2 (t ) + 16 y (t ) ;

xˆ1 (0 ) = 0 xˆ2 (0 ) = 0

(4.3.23) (4.3.24)

The analog computer simulation diagram for the observer-based state feedback controller is shown in Figure 4.6. The controller transfer function is H (s ) = −k (sI − ( A − bk − Lc))−1 L =

76 s + 256 s 2 + 6 s + 25

(4.3.25)

4.4 REDUCED-ORDER OBSERVER Because p outputs y(t) are linear combinations of n states x(t), output equations y (t ) = Cx (t ) represent p equations in n unknowns where p < n . Therefore, there is no need to estimate all the states, and the order of the observer dynamics can be reduced to be n–p. Here, as an example, a single-output system (Kailath, 1980) is considered. Let c = ⎡⎣0

0

.

.

0

1⎤⎦

(4.4.1)


192


Then, y = x n ; i.e., x n is directly being measured. Therefore, we have only to observe or estimate x r = ⎡⎣ x1

x2

.

.

x n −1 ⎤⎦

T

(4.4.2)

Now, the state space equation is partitioned as ⎡ xr ⎤ ⎡ Ar ⎢ ⎥=⎢ ⎣ x n ⎦ ⎣ cr

br ⎤ ⎡ x r ⎤ ⎡ g r ⎤ ⎥ ⎢ ⎥ + ⎢ ⎥u ann ⎦ ⎣ x n ⎦ ⎣ gn ⎦

(4.4.3)

or xr = Ar x r + b r y + g r u

(4.4.4)

x n = cr x r + ann y + gn u

(4.4.5)

yr = c r x r

(4.4.6)

Define

Then, Equation 4.4.5 can be written as yr = y − ann y − gn u

(4.4.7)

Because the right-hand side of Equation 4.4.7 is known, yr can be viewed as the output, and cr as the output vector. Also, from Equation 4.4.4, b r y + g r u can be considered as the input for the x r dynamics. Analogous to the equation for the fullorder observer dynamics, the reduced-order observer can be now described by xˆ r = Ar xˆ r + b r y + g r u + r ( yr − cr xˆ r )

(4.4.8)

where r is an ( n − 1) dimensional vector which has to be determined. Define the error in the estimated state vector as x r = x r − xˆ r Subtracting (4.4.9) from (4.4.4),

(4.4.9)



193

d dt

br y (t )

r

xˆ r

xˆ r

r a nn

xn ( t )

gr

u(t )

r gn

Ar

rc r

FIGURE 4.7 Reduced-order observer with a differentiator.

r

br y (t )

r a nn

q

q

xˆ r

xn ( t )

u(t )

gr

r gn

Ar

rcr

FIGURE 4.8 Reduced-order observer without a differentiator.

x r = ( Ar − r cr ) x r

(4.4.10)

It has been shown that {c r , Ar } is observable, provided {c, A} is observable (Kailath, 1980). Therefore, r can be calculated for any choice of the desired characteristic polynomial for (4.4.10) when the state space realization is observable. We can then guarantee that x r (t ) → 0 as t → ∞ irrespective of the value of x r (0 ) . The problem of observing unmeasured states has been solved only in theory because the knowledge of yr via (4.4.7) implies differentiating the output signal to obtain y. The resulting system is shown in Figure 4.7. From (4.4.7) and (4.4.8), xˆ r = ( Ar − r cr ) xˆ r + b r y + g r u + r ( y − ann y − gn u ) = ( Ar − r cr ) xˆr + (b r − r ann ) y + (g r − r gn )u + r y

(4.4.11)

In order to get rid of the differentiator, the system in Figure 4.7 is modified by changing the path with the block r from the input to the output side of the integrator (Figure 4.8). The output of the integrator is now q(t ) = xˆ r (t ) − r y (t )

(4.4.12)


194


x(t )

f (t ) m

FIGURE 4.9 A simple mass.

From (4.4.11) and (4.4.12), q = ( Ar − r cr )q(t ) + ( Ar r − r cr r + b r − r ann ) y (t ) + (g r − r gn )u (t ) (4.4.13) With y (t ) and u (t ) known, Equation 4.4.13 has to be solved in real time to determine q(t). Then the states are estimated by xˆ r (t ) = q(t ) + r y (t )

(4.4.14)

EXAMPLE 4.4: A SIMPLE MASS Consider a simple mass m which is subjected to a control force f (t ) (Figure 4.9). The differential equation of motion is x = u

(4.4.15)

where u (t ) =

f (t ) m

(4.4.16)

Define states as x1 (t ) = x

and

x2 = x

(4.4.17)

Then, the state equations are ⎡ x1 ⎤ ⎡0 ⎢ ⎥=⎢ ⎣ x2 ⎦ ⎣ 1

0 ⎤ ⎡ x1 ⎤ ⎡ 1 ⎤ ⎥ ⎢ ⎥ + ⎢ ⎥u 0 ⎦ ⎣ x2 ⎦ ⎣ 0 ⎦

(4.4.18)

Let us assume that we have only one sensor that can measure the position of the mass. Then, the output equation is



195

y (t ) = ⎡⎣0

⎡x ⎤ 1⎤⎦ ⎢ 1 ⎥ ⎣ x2 ⎦

(4.4.19)

The objective is to design a reduced-order observer to estimate the velocity of the mass. This is interesting because the velocity will be estimated from the position of the mass via an integrator, i.e., without using a differentiator. In view of the partitioning shown in Equation 4.4.3, ⎡0 A=⎢ ⎣1

0 ⎤ ⎡ Ar ⎥=⎢ 0 ⎦ ⎣ cr

br ⎤ ⎥ ann ⎦

⎡1⎤ ⎡ g ⎤ b=⎢ ⎥=⎢ r⎥ ⎣0 ⎦ ⎣ gn ⎦

(4.4.20)

(4.4.21)

Hence, the state error dynamics is written as x1 = ( Ar − lr cr ) x1 = −lr x1

(4.4.22)

x1 (t ) = x1 (0 )e − lr t

(4.4.23)

q = −lr q − lr2 y + u = −lr (q + lr y ) + u = −lr xˆ1 + u

(4.4.24)

xˆ1 (t ) = q (t ) + lr y (t )

(4.4.25)

Solving (4.4.22),

Equation 4.4.13 is written as

where

The simulation diagram for the velocity estimator is shown in Figure 4.10.

4.5 RESPONSE OF A LINEAR CONTINUOUS-TIME SYSTEM TO WHITE NOISE Consider the following linear system: x = Ax (t ) + Bw(t )

(4.5.1)


196


y(t )

x2 ( t )

x(t )

r

u(t )

q

q xˆ1

xˆ (t )

r

FIGURE 4.10 A reduced-order observer to estimate velocity.

where w(t ) is a zero-mean white noise vector (Appendix D) with the intensity Q; i.e., E ( w(t )) = 0

(4.5.2)

E [ w(t ) w T (t + τ)] = Q δ(τ)

(4.5.3)

Initial condition x(0 ) is assumed to be a random variable vector independent of w(t ); i.e., E [ x (0 ) w T (τ)] = E [ x (0 )]E [ w T (τ)] = 0

(4.5.4a)

E( x (0 )) = x 0

(4.5.4b)

E [( x (0 ) x T (0 )] = P0

(4.5.5)

and

The solution of (4.5.1) is t

x (t ) = e At x (0 ) +

∫e

A (t − τ )

Bw(τ) d τ

0

Mean of x(t) ⎛ E (x(t )) = E (e At x(0 )) + E ⎜ ⎜ ⎝

t

∫ 0

⎞ e A (t − τ ) Bw( τ )d τ ⎟ ⎟ ⎠

(4.5.6)



197

t

= e E ( x (0 )) + At

∫e

A (t − τ )

BE ( w(τ)) d τ = e At x 0

(4.5.7)

0

Variance of x(t) Define the covariance matrix: P (t ) = E ( x (t ) x T (t )]

(4.5.8)

Using (4.5.6), t

∫ x(0)w (τ)B e

T

x (t ) x T (t ) = e At x (0 ) x T (0 )e A t + e At

T

T

AT ( t − τ )

dτ

0

t

t

∫

T

+ e A (t − τ ) Bw(τ) x T (0 )e A τ d τ + 0

(4.5.9)

t

∫ ∫e 0

A (t − τ )

Bw(τ) w T ( ν) BT e A

T

(t − ν)

d τd ν

0

Taking expected values on both sides, t

P (t ) = e E ( x (0 ) x (0 ))e At

T

AT t

+

t

∫ ∫e 0

A (t − τ )

BQ δ( ν − τ) BT e A

T

(t − ν)

d νd τ

(4.5.10)

0

t

T

P (t ) = e At P (0 )e A t +

∫

e A (t − τ ) BQBT e A

T

(t − τ )

dτ

(4.5.11)

0

Differentiating (4.5.11) with respect to t, dP = AP + PAT + BQBT ; dt

P (0 ) = P0

(4.5.12)

For a stable system, dP →0 dt

as

t→∞

(4.5.13)

Therefore, in steady state, the covariance matrix P is governed by the Lyapunov equation (Chapter 2, Section 2.15):


198


AP + PAT + BQBT = 0 EXAMPLE 4.5: A SPRING-MASS DAMPER SUBJECTED

(4.5.14)

TO A

STOCHASTIC FORCE

Consider the spring-mass-damper system (Chapter 2, Figure 2.4) where the force f(t) is a zero-mean Gaussian white noise with the following statistics: E ( f (t ) f (t + τ)) = 0.1δ(τ)

(4.5.15)

Let the system have the following parameters: m = 1 kg, α = 0.2 N − s/m , and β = 1N /m . Then, ⎡0 A=⎢ ⎣ −1

1 ⎤ ⎥ −0.2 ⎦

and

⎡0 BQBT = 0.1 ⎢ ⎣0

0⎤ ⎥ 1⎦

(4.5.16a,b)

Solving Equation 4.5.14 via the Matlab routine “lyap,” ⎡0.25 P=⎢ ⎣ 0

0 ⎤ ⎥ 0.25 ⎦

(4.5.17)

The response (Figure 4.11) is also generated by numerical solution of the state equations (2.5.16) in Chapter 2. The Matlab program is provided (Program 4.1). Note that the Gaussian white noise is generated by the method presented in Appendix D. Numerical results provide the following value of P: ⎡ 0.2442 Psim = ⎢ ⎣ 4.53e − 5

4.53e − 5 ⎤ ⎥ 0.2516 ⎦

which is close to the analytical solution. MATLAB PROGRAM 4.1: LINEAR SYSTEM SUBJECTED % clear all close all global wC % wone=randn(1,10000); w=sqrt(2)*wone;

TO

WHITE NOISE

(4.5.18)



199

2

1.5

1

x(t)

0.5

0

–0.5

–1

–1.5

0

50

100

150

200

250 300 time (sec)

350

400

450

500

FIGURE 4.11 Response obtained from Matlab Program 4.1.

% % disp(1)=0; Vel(1)=0; tv(1)=0; tinit=0; delt=0.05; % %PSD of White Noise=2*delt=0.1; for i=2:10000 tf=tinit+delt; wC=w(1,i-1); [t,Y]=ode45(@ex4p6,[tinit tf],[disp(i-1) Vel(i-1)]); yud=flipud(Y); tv(i)=tf;


200


disp(i)=yud(1,1); Vel(i)=yud(1,2); tinit=tf; end ---------------------------------------function dx=ex4p6(t,x) % global wC dx=zeros(2,1); dx(1)=x(2); dx(2)=-0.2*x(2)-x(1)+wC;

4.6 KALMAN FILTER: OPTIMAL STATE ESTIMATION The Kalman filter deals with optimal selection of observer poles, and is dual to the linear quadratic control presented in Chapter 3.

STATE DYNAMICS x (t ) = Ax (t ) + Bu(t ) + w(t )

(4.6.1)

The vector w(t ) is a stochastic process called process (or plant) noise. It is assumed that w(t ) is a continuous-time Gaussian white noise vector (Appendix D). Its mathematical characterization is E ( w(t )) = 0

(4.6.2)

E ( w(t ) w T (t + τ)) = W δ(τ)

(4.6.3)

and

The matrix W is called the intensity matrix with the property W = W T > 0 .

MEASUREMENT EQUATION y(t ) = Cx(t ) + θ(t )

(4.6.4)



201

where y(t ) is the sensor noise vector and θ(t ) is a continuous-time Gaussian white noise vector. It is assumed that E (θ(t )) = 0

(4.6.5)

E (θ(t )θT (t + τ )) = Θδ( τ )

(4.6.6)

and

where Θ = ΘT > 0 . It is also assumed that process and measurement noise vectors are uncorrelated; i.e., E (θ(t )wT (t + τ )) = E (θ(t ))E (wT (t + τ )) = 0

(4.6.7)

OBSERVER DYNAMICS From Equation 4.2.3, xˆ = Axˆ + Bu(t ) + L ( y (t ) − Cxˆ (t ))

(4.6.8)

Define the state error vector as x = x (t ) − xˆ (t )

(4.6.9)

x = ( A − LC )x + ψ(t )

(4.6.10)

ψ(t ) = w(t ) − Lθ(t )

(4.6.11)

E ( ψ(t )) = E (w(t )) − LE (θ(t )) = 0

(4.6.12)

Subtracting (4.6.8) from (4.6.1),

where

Consider ψ(t )ψT (t + τ ) = [ w(t ) − Lθ(t )][ w(t + τ ) − Lθ(t + τ )]T = w((t )wT (t + τ ) − Lθ(t )wT (t + τ ) − w(t )θT (t + τ )LT + Lθ(t )θT (t + τ )LT Using (4.6.3), (4.6.6), and (4.6.7),

(4.6.13)


202


E ( ψ(t )ψT (t + τ )) = (W + LΘLT )δ( τ )

(4.6.14)

The state error dynamics (Equation 4.6.10) is represented by a linear system subjected to white noise vector ψ(t ) with the intensity described by (4.6.14). Assuming that the matrix ( A − LC ) is stable, E ( x (t )) → 0

as

t→∞

(4.6.15)

and ( A − LC ) Σ + Σ( A − LC )T + W + L ΘLT = 0

(4.6.16)

where Σ is the steady state error covariance matrix defined as Σ = lim E (x (t )x T (t ))

(4.6.17)

J = tr[ Σ] = E ( x12 (t )) + E ( x 22 (t )) + … + E ( x n2 (t ))

(4.6.18)

t →∞

Problem Find L such that

is minimized subject to constraints (4.6.16). Solution Introducing symmetric matrix P as the Lagrange multiplier, g ( Σ, L, P ) = tr ( Σ) + tr (U ( Σ, L ) P )

(4.6.19)

where U ( Σ, L ) is the left-hand side of Equation 4.6.16. As a result, g( Σ, L, P ) = tr ( Σ) + tr (( A − LC )ΣP ) + tr ( Σ( A − LC )T P ) + tr (WP ) + tr ( LΘLT P )

(4.6.20)

Using the Lemma for the optimal solution (Appendix E), ∂g = I + ( A − LC )T P + P ( A − LC ) = 0 ∂Σ

(4.6.21)



203

∂g = − P ΣC T − P ΣC T + PL Θ + PL Θ = 2 P (− ΣC T + L Θ) = 0 ∂L

(4.6.22)

∂g = Σ( A − LC )T + ( A − LC ) Σ + W + L ΘLT = 0 ∂P

(4.6.23)

Assuming that ( A − LC ) is stable, P > 0 from the well-known result of the solution of the Lyapunov equation (Chapter 2, Section 2.15). Hence, from Equation 4.6.22, − ΣC T + L Θ = 0

(4.6.24)

L = ΣC T Θ −1

(4.6.25)

ΣAT + AΣ − ΣC T Θ −1C Σ + W = 0

(4.6.26)

or


Equation 4.6.26 is called the filter algebraic Riccati equation (FARE). The important statistical properties of the Kalman filter (Anderson and Moore, 1990) are as follows: 1. The state error vector is orthogonal to estimate the state vector in the following sense: E ( x xˆ T ) = 0

(4.6.27)

v (t ) = y (t ) − Cxˆ (t )

(4.6.28)

2. The signal

is white noise with zero mean, and is called the innovation process. Furthermore, E (v (t )v T (t + τ)) = Θδ(τ)

(4.6.29)


204


EXAMPLE 4.6: A SIMPLE MASS (FIGURE 4.9) OR A DOUBLE INTEGRATOR SYSTEM WITH STOCHASTIC DISTURBANCES Consider the problem of finding an optimal state estimator for the system ⎡ x1 ⎤ ⎡0 ⎢ ⎥=⎢ ⎣ x2 ⎦ ⎣0

⎡0 ⎤ 1 ⎤ ⎡ x1 ⎤ ⎡0 ⎤ ⎥ ⎢ ⎥ + ⎢ ⎥ u (t ) + ⎢ ⎥ g (t ) 0 ⎦ ⎣ x2 ⎦ ⎣ 1 ⎦ ⎣1⎦

(4.6.30)

⎡ x (t ) ⎤ 0 ⎤⎦ ⎢ 1 ⎥ + θ(t ) ⎣ x2 (t ) ⎦

(4.6.31)

and

y (t ) = ⎡⎣1 Let

E ( g (t ) g (t − τ)) = δ(t − τ)

(4.6.32)

E (θ(t )θ(t − τ)) = ρδ(t − τ)

(4.6.33)

and

Therefore, ⎡0 W =⎢ ⎣0

0⎤ ⎥ 1⎦

and

Θ=ρ

(4.6.34)

Define ⎡Σ Σ = ⎢ 11 ⎣ Σ12

Σ12 ⎤ ⎥ Σ 22 ⎦

(4.6.35)

Then, ⎡Σ AΣ = ⎢ 12 ⎣ 0

ΣC T

Σ 22 ⎤ ⎥; 0 ⎦

⎡Σ ΣAT = ( AΣ)T = ⎢ 12 ⎣ Σ 22

1 1 ⎡ Σ2 C Σ = ⎢ 11 ρ ρ ⎣ Σ11Σ12

Σ11Σ12 ⎤ ⎥ 2 Σ12 ⎦

0⎤ ⎥ 0⎦

(4.6.36a,b)

(4.6.37)



205

Equation 4.6.26 yields the following three algebraic equations:

2 Σ12 −

Σ 22 −

1−

2 Σ11 =0 ρ

(4.6.38)

Σ11Σ12 =0 ρ

(4.6.39)

2 Σ12 =0 ρ

(4.6.40)

Solving these three equations simultaneously, ⎡ 2 ρ0.75 Σ=⎢ ⎢⎣ ρ

ρ ⎤ ⎥>0 2 ρ0.25 ⎥⎦

(4.6.41)

From (4.6.25), ⎡ 2 / ρ0.25 ⎤ ⎥ L=⎢ ⎢⎣ 1 / ρ ⎥⎦

(4.6.42)

For ρ = 0.01, Equation 4.6.41 yields ⎡0.0447 Σ=⎢ ⎣ 0.1

0.1 ⎤ ⎥ 0.4472 ⎦

(4.6.43)

The errors in states from numerical simulations are presented in Figure 4.12 and Figure 4.13. Also, ⎡ 0.0461 Σ sim = ⎢ ⎣0.1012 MATLAB PROGRAM 4.2: KALMAN FILTER % clear all close all

0.1012 ⎤ ⎥ 0.4603 ⎦

(4.6.44)


206


1 ρ = 0.01 0.8

ΔT = 0.1

Error in Estimate of x1

0.6 0.4 0.2 0

–0.2 –0.4 –0.6 0

5

10

15

20

25 30 time (sec)

35

40

45

50

45

50

FIGURE 4.12 Error x1 (t ) from the Kalman filter (Matlab Program 4.2). 2 ρ = 0.01 1.5 ΔT = 0.1

Error in Estimate of x2

1 0.5 0 –0.5 –1 –1.5 –2 –2.5

0

5

10

15

20

25 30 time (sec)

35

40

FIGURE 4.13 Error x 2 (t ) from the Kalman filter (Matlab Program 4.2).



global wthC Aobs % rho=0.01; SIGMA=[sqrt(2)*(rho^0.75) sqrt(rho);sqrt(rho) sqrt(2)*(rho^0.25)] delt=0.1; % load rand_chap4 w=wone/sqrt(delt); load th_rand_chap4 th=sqrt(rho)*thone/sqrt(delt); % A=[0 1;0 0]; L=[sqrt(2)/(rho^0.25) sqrt(1/rho)]'; C=[1 0]; Aobs=A-L*C; % % disp_er(1)=1; Vel_er(1)=0.0; tv(1)=0; tinit=0; % for i=2:5000 tf=tinit+delt; wC=w(i-1); thC=th(i-1); wthC=[0;wC]-L*thC; [t,Y]=ode45(@ex4p6KF,[tinit tf],[disp_er(i-1) Vel_er(i-1)]); yud=flipud(Y);

207


208


tv(i)=tf; disp_er(i)=yud(1,1); Vel_er(i)=yud(1,2); tinit=tf; end % ---------------------------------------function dx=ex4p6KF(t,x) % global wthC Aobs % dx=zeros(2,1); dx=Aobs*x+wthC;

4.7 STOCHASTIC OPTIMAL REGULATOR IN STEADY STATE It is assumed that all states are known. However, the plant dynamics is subjected to stochastic disturbances.


(4.7.1)


(4.7.2)

E ( w(t ) w T (t + τ)) = W δ(τ)

(4.7.3)

and

The matrix W is called the intensity matrix with the property W = W T > 0.



209

OBJECTIVE FUNCTION Minimize ⎧⎪ 1 J = lim E ⎨ t f →∞ ⎪⎩ t f

tf

∫ 0

⎫⎪ [ xT (t )Qx(t ) + uT (t )Ru(t )dt ]⎬ ⎪⎭

(4.7.4)

Let the optimal control be u(t ) = − K x (t )

(4.7.5)

x (t ) = ( A − BK ) x (t ) + w(t )

(4.7.6)

Then, from (4.7.1),

Assuming that the matrix ( A − BK ) is stable, E ( x (t )) → 0

as

t→∞

(4.7.7)

and ( A − BK ) M + M ( A − BK )T + W = 0

(4.7.8)

where M is the steady state covariance matrix defined as M = lim E (x(t )xT (t )) t →∞

(4.7.9)

Assuming the asymptotic stability of the closed-loop system, Equation 4.7.4 can also be written as J = lim E {xT (t )Qx(t ) + uT (t ) Ru(t )} t →∞

(4.7.10)

Equation 4.7.10 can be rewritten as J = lim E {tr (x(t )xT (t )Q + u(t )uT (t )R)} t →∞

Using (4.7.5) and interchanging E(.) and tr operations,

(4.7.11)


210


J = tr ( MQ + KMK T R )

(4.7.12)

Problem Find K such that J Equation 4.7.12 is minimized subject to constraints (4.7.8). Solution First, Equation 4.7.12 is rewritten as J = tr ( MQ + RKMK T )

(4.7.13)

Introducing a symmetric matrix P as the Lagrange multiplier, g ( M , K , P ) = tr ( MQ + RKMK T ) + tr (U ( M , K ) P )

(4.7.14)

where U ( M , K ) is the left-hand side of Equation 4.7.8. As a result, g( M , K , P ) = tr ( MQ + RKMK T ) + tr (( A − BK )MP ) + tr ( M ( A − BK )T P ) + tr (WP )

(4.7.15)

Using the Lemma for the optimal solution (Appendix E), ∂g = M ( A − BK )T + ( A − BK ) M + W = 0 ∂P

(4.7.16)

∂g = RKM + RKM − BT PM − BT PM = 2 ( RK − BT P ) M = 0 ∂K

(4.7.17)

∂g = Q + K T RK + ( A − BK )T P + P ( A − BK ) = 0 ∂M

(4.7.18)

For a stable ( A − BK ) , M > 0 from the well-known result of the solution of the Lyapunov equation (Chapter 2, Section 2.15). Hence, from Equation 4.7.17, K = R −1BT P

(4.7.19)

Substituting (4.7.19) into (4.7.18), the Riccati equation is obtained: PA + AT P − PBR −1BT P + Q = 0

(4.7.20)



211

To find the optimal value of the objective function, Equation 4.7.13 is written as J = tr ( MQ + MK T RK ) = tr ( M (Q + K T RK ))

(4.7.21)

Then, using (4.7.18) and (4.7.16), J = tr (WP ) = tr ( PW )

(4.7.22)

u = −K x (t )

(4.7.23)

K = R −1BT P

(4.7.24)

PA + AT P − PBR −1BT P + Q = 0

(4.7.25)

Summary The optimal control law is

The minimum value of J is J = tr ( PW )

(4.7.26)

EXAMPLE 4.7: ACTIVE SUSPENSION SYSTEM Consider that the road profile, x0 (t ), is a white noise (Chapter 3, Example 3.11) with the following statistics: E ( x0 (t ) x0 (t + τ)) = νδ(τ)

(4.7.27)

In this case, the state equations are written as

⎡ 0 ⎡ x1 ⎤ ⎢ ⎢ ⎥ ⎢ 0 ⎢ x2 ⎥ = ⎢ λ ⎢ x3 ⎥ ⎢ − 1 ⎢ ⎥ ⎢ m1 ⎢⎣ x 4 ⎥⎦ ⎢ ⎣ 0 where

0 0

1 0

0

0

0

0

⎡ 0 ⎤ 0⎤ ⎢ ⎥ ⎡ ⎤ x 0 ⎥ ⎥ 1 1⎥ ⎢ ⎥ ⎢ ⎢ 1 ⎥ x ⎥ ⎢ 2 ⎥ + ⎢− ⎥ u (t ) + w(t ) ⎢ ⎥ 0 ⎥ x3 m1 ⎥ ⎢ ⎥⎢ ⎥ ⎢ 1 ⎥ ⎢x ⎥ ⎥ 0 ⎥⎦ ⎣ 4 ⎦ ⎢ ⎢⎣ m2 ⎥⎦

(4.7.28)


212


w(t ) = b w x0 (t )

b Tw = [0

0

λ1 m1

(4.7.29a)

0]

(4.7.29b)

From (4.7.27) and (4.7.29), W = b w b Tw ν2

(4.7.30)

u (t ) = −kx

(4.7.31)

Then the controller

where k computed by (3.5.84) minimizes the following objective function: J = lim E {xT (t )Qx(t ) + ρu 2 } t →∞

(4.7.32)

From (4.7.26) and (4.7.30), the minimum value of the objective function will be

J = (P11 + P22 + P33 + P44 ) ν2

EXAMPLE 4.8: A SIMPLE MASS

OR A

λ12 m12

(4.7.33)

DOUBLE INTEGRATOR SYSTEM

Consider the system (4.6.30) again. Weighting matrices Q and R are chosen as follows: ⎡1 Q=⎢ ⎣0

0⎤ ⎥ 0⎦

and

p12 ⎤ ⎥ p22 ⎦

with

R = ρc

(4.7.34a,b)

Define ⎡p P = ⎢ 11 ⎣ p21

p12 = p21

(4.7.35)

From Equation 4.7.25, the following nonlinear algebraic equations are obtained:



213

−

2 p12 +1= 0 ρc

2 p12 −

p11 −

2 p22 =0 ρc

p12 p22 =0 ρc

(4.7.36a)

(4.7.36b)

(4.7.36c)

Solving Equations 4.7.36a to Equation 4.7.36c simultaneously, ⎡ 2 ρc P=⎢ ⎢⎣ ρc

⎤ ⎥ 2 0.5 ρc0.75 ⎥⎦

(4.7.37)

K = [ ρc

2 0.5 ρc0.75 ]

(4.7.38)

⎡0 PW = ⎢ ⎢⎣0

ρc ⎤ ⎥ 2 ρ0c .75 ⎥⎦

(4.7.39)

J = tr ( PW ) = 2 0.5 ρc0.75

(4.7.40)

ρc

From (4.7.24),

From (4.7.37) and (4.6.34),

0.5

From (4.7.26),

For ρc = 0.05 , Equation 4.7.40 yields J = 0.1495 From numerical simulation (Matlab Program 4.3), Jsim = 0.1581 The response x1 (t ) is shown in Figure 4.14.


214


1 ρc = 0.05

0.8

ΔT = 0.1 0.6

Displacement, x2(t)

0.4 0.2 0 –0.2 –0.4 –0.6 –0.8

0

20

40

60

80

100 120 time (sec)

140

160

180

FIGURE 4.14 Displacement x1 (t ) from stochastic regulator (Matlab Program 4.3).

MATLAB PROGRAM 4.3: STOCHASTIC REGULATOR % clear all close all global wC Acl % rhoc=0.05; delt=0.1; % load rand_chap4 w=wone/sqrt(delt); % A=[0 1;0 0]; K=[sqrt(1/rhoc) sqrt(2)/(rhoc^0.25)]; B=[0 1]';

200



215

Acl=A-B*K; % J=sqrt(2)*(rhoc)^0.75 % disp(1)=1; Vel(1)=0.0; tv(1)=0; tinit=0; % for i=2:5000 tf=tinit+delt; wC=w(i-1); [t,Y]=ode45(@ex4stochR,[tinit tf],[disp(i-1) Vel(i1)]); yud=flipud(Y); tv(i)=tf; disp(i)=yud(1,1); Vel(i)=yud(1,2); uin(i)=-K(1)*disp(i)-K(2)*Vel(i); tinit=tf; end -----------------------------------------function dx=ex4stochR(t,x) % global wC Acl % dx=zeros(2,1); dx=Acl*x+[0;1]*wC;

4.8 LINEAR QUADRATIC GAUSSIAN (LQG) CONTROL It is assumed that all states are not known. The plant dynamics and output measurements are subjected to stochastic disturbances.


216



(4.8.1)

The vector w(t ) is a stochastic process called process (or plant) noise. It is assumed that w(t ) is a continuous-time Gaussian white noise vector. Its mathematical characterization is E ( w(t )) = 0

(4.8.2)

E ( w(t ) w T (t + τ)) = W δ(τ)

(4.8.3)

and

The matrix W is called the intensity matrix with the property W = W T ≥ 0.

MEASUREMENT EQUATION y(t ) = Cx(t ) + θ(t )

(4.8.4)

where y(t ) is the sensor noise vector and θ(t ) is a continuous-time Gaussian white noise vector. It is assumed that E (θ(t )) = 0

(4.8.5)

E (θ(t )θT (t + τ )) = Θδ( τ )

(4.8.6)

and

where Θ = ΘT > 0. It is also assumed that process and measurement noise vectors are uncorrelated; i.e., E (θ(t )wT (t + τ )) = E (θ(t ))E (wT (t + τ )) = 0

(4.8.7)

OBJECTIVE FUNCTION Minimize ⎧ ⎪1 J = lim E ⎨ t f →∞ ⎪⎩ t f

tf

∫ 0

⎫ ⎪ [ xT (t )Qx(t ) + uT (t )Ru(t )dt ]⎬ ⎪⎭

(4.8.8)



217

Assuming the asymptotic stability of the closed-loop system, Equation 4.8.8 can also be written as J = lim E {xT (t )Qx(t ) + uT (t ) Ru(t )} t →∞

(4.8.9)

Let the states be estimated by the Kalman filter (4.6.8) with the observer gain L given by (4.6.25). Substituting x = xˆ + x , T Q ) + 2tr ( x xˆ T Q ) (4.8.10) x T Qx = xˆ T Qxˆ + x T Qx + 2 xˆ T Qx = xˆ T Qxˆ + tr ( xx Substituting (4.8.10) into (4.8.19), interchanging tr and E(.) operations, and using the orthogonality property (4.6.27), J = lim E {xˆ T (t )Qxˆ (t ) + uT (t )Ru(t )} + tr ( ΣQ )

(4.8.11)

xˆ = Axˆ + Bu(t ) + Lv (t )

(4.8.12)

t →∞

where

and v(t ) is a white noise defined by (4.6.28). Because the second term tr ( ΣQ ) in (4.8.11) is a constant, the minimization of J (Equation 4.8.11) subject to constraints (4.8.12) is the stochastic regulator problem solved in Section 4.6. Therefore, the optimal control law is u(t ) = − K xˆ (t )

(4.8.13)

where K is defined by (4.7.24) and (4.7.25). In summary, the optimal control law is obtained by separately solving the optimal estimation problem and the deterministic LQ control problem. This is known as the separation theorem. The LQG control system is developed by completing the following steps (Athans, 1992): Step I: The optimal state estimator xˆ = Axˆ + Bu(t ) + L ( y (t ) − Cxˆ )

(4.8.14)

L = ΣC T Θ −1

(4.8.15)

where


218


and AΣ + ΣAT + W − ΣC T Θ −1C Σ = 0

(4.8.16)

Equation 4.8.16 is known as the filter algebraic Riccati equation (FARE). Step II: The optimal control law is u(t ) = − K xˆ (t )

(4.8.17)

K = R −1BT P

(4.8.18)

PA + AT P − PBR −1BT P + Q = 0

(4.8.19)

where

and

Equation 4.8.19 is known as the control algebraic Riccati equation (CARE).

CLOSED-LOOP SYSTEM DYNAMICS Define the state error vector as x = x − xˆ

(4.8.20)

Substituting (4.8.18) and (4.8.20) into (4.8.1), x = ( A − BK ) x (t ) + BK x (t ) + w(t )

(4.8.21)

Subtracting (4.8.14) from (4.8.1), x = ( A − LC )x − Lθ(t ) + w(t )

(4.8.22)

Putting (4.8.21) and (4.8.22) in the matrix form, ⎡x ⎤ ⎡ A − BK ⎢ ⎥=⎢ ⎢⎣x ⎥⎦ ⎣ 0 Equation 4.8.23 is equivalent to

BK ⎤⎡x⎤ ⎡ w(tt ) ⎤ ⎥⎢ ⎥ + ⎢ ⎥ A − LC ⎦⎣x ⎦ ⎣w(t ) − Lθ(t )⎦

(4.8.23)



219

⎡x ⎤ ⎡ A ⎢ ⎥=⎢ ⎢⎣xˆ ⎥⎦ ⎣ LC

MINIMUM VALUE

OF THE

⎤⎡x⎤ ⎡ w(t ) ⎤ − BK ⎥⎢ ⎥ + ⎢ ⎥ A − BK − LC ⎦⎣xˆ ⎦ ⎣ Lθ(t )⎦

(4.8.24)

OBJECTIVE FUNCTION

Drawing analogy with the stochastic regulator problem, lim E {xˆ T (t )Qxˆ (t ) + uT (t )Ru(t )} = tr ( PLΘLT )

(4.8.25)

E ( Lv (t )v T (t + τ) LT ) = L ΘLT δ(τ)

(4.8.26)

t →∞

where

Therefore, using (4.8.11), the minimum value of the objective function is given by

{

} { }

J = tr PL ΘLT + tr ΣQ

(4.8.27)

EXAMPLE 4.9: A SIMPLE MASS OR A DOUBLE INTEGRATOR SYSTEM (LQG CONTROL) Consider the system (4.6.30) again. The weighting matrices Q and R in (4.8.8) are also defined by Equation 4.7.34a and Equation 4.7.34b. Then, from (4.8.27), J = ρ[21 2 ρc + 2 1 2 ρc + 2 22 ρc0.75 ] + 2 ρ0.75

(4.8.28)

where from Equation 4.6.42,

1 =

2 ρ0.25

and

2 =

1 ρ

(4.8.29)

Substituting (4.8.29) into (4.8.28), J = 2 2 ρc ( ρ + ρ0.25 ) + 2 (ρc0.75 + ρ0.75 )

(4.8.30)

For ρ = 0.01 and ρc = 0.05 , J = 0.5357 From numerical simulations (Matlab Program 4.4),

(4.8.31)


220


1.5

ρc = 0.05

ρ = 0.01

ΔT = 0.1

1

0.5

x1(t)

0

–0.5

–1

–1.5

–2

0

20

40

60

80

100 120 time (sec)

140

160

180

200

FIGURE 4.15 Displacement x1 (t ) from the linear quadratic Gaussian (LQG) control (Matlab Program 4.4).

Jsim = 0.5529 The response x1 (t ) is shown in Figure 4.15. MATLAB PROGRAM 4.4: LQG CONTROL % clear all close all global wC thC K L % rho=0.01; rhoc=0.05; delt=0.1; % load rand_chap4

(4.8.32)



221

w=wone/sqrt(delt); load th_rand_chap4 th=sqrt(rho)*thone/sqrt(delt); % A=[0 1;0 0]; B=[0 1]'; C=[1 0]; % L=[sqrt(2)/(rho^0.25) sqrt(1/rho)]'; K=[sqrt(1/rhoc) sqrt(2)/(rhoc^0.25)]; % J=2*sqrt(2*rhoc)*((rhoc)^0.5+(rho)^0.25)+sqrt(2)*((r hoc)^0.75+(rho)^0.75) % disp(1)=1; Vel(1)=0.0; disph(1)=0; Velh(1)=0; tv(1)=0; tinit=0; % for i=2:5000 tf=tinit+delt; wC=w(i-1); thC=th(i-1); [t,Y]=ode45(@ex4p6lqg,[tinit tf],[disp(i-1) Vel(i-1) disph(i-1) Velh(i-1)]); yud=flipud(Y); tv(i)=tf; disp(i)=yud(1,1); Vel(i)=yud(1,2);


222


disph(i)=yud(1,3); Velh(i)=yud(1,4); uin(i)=-K(1)*disph(i)-K(2)*Velh(i); tinit=tf; end % function dx=ex4p6lqg(t,x) % global wC thC K L % dx=zeros(4,1); u=-K(1)*x(3)-K(2)*x(4); dx(1)=x(2); dx(2)=u+wC; dx(3)=x(4)+L(1)*(x(1)+thC-x(3)); dx(4)=u+L(2)*(x(1)+thC-x(3));

4.9 IMPACT OF MODELING ERRORS ON OBSERVER-BASED CONTROL 4.9.1 STRUCTURED PARAMETRIC UNCERTAINTIES Consider the state space model of the plant: x = Ax (t ) + Bu(t )

(4.9.1)

y (t ) = Cx (t )

(4.9.2)

The elements of matrices A, B, and C are assumed to be uncertain. This type of modeling error is known as a structured parametric error because the structure of the model is considered accurate, only the parameters of the model are not well known. Let Aˆ , Bˆ , and Cˆ be estimates of A, B, and C, respectively. Then, the observer-based controller can be defined as u = −K xˆ

(4.9.3)



223

xˆ = Aˆ xˆ + Bˆ u + L ( y − Cˆxˆ )

(4.9.4)

Combining (4.9.1) through (4.9.4), the closed-loop system dynamics is defined as ⎡ x ⎤ ⎡x ⎤ ⎢ ⎥ = Asysp ⎢ ⎥ ⎣ xˆ ⎦ ⎣ xˆ ⎦

(4.9.5)

where ⎡ A Asysp = ⎢ ⎢⎣ LC

⎤ − BK ˆ − LCˆ ⎥⎥ Aˆ − BK ⎦

(4.9.6)

ˆ ) and ( Aˆ − LCˆ ). In fact, it is Now, eigenvalues of Asysp are not those of ( Aˆ − BK possible that the closed-loop system is unstable under parametric errors. EXAMPLE 4.10: BENCHMARK ROBUST CONTROL PROBLEM (WIE AND BERNSTEIN, 1992) The differential equations of motion for the system in Figure 4.16 are m1q1 + k (q1 − q2 ) = u

(4.9.7)

m2 q2 + k (q2 − q1 ) = 0

(4.9.8)

x1 = q1 , x2 = q2 , x3 = q1 , and x 4 = q2

(4.9.9)

Defining states as

The state space model is x = Ax + Bu

(4.9.10)

y = Cx

(4.9.11)

q1 u

q2 k

m1

FIGURE 4.16 A two-mass system.

m2

y


224


where ⎡ 0 ⎢ 0 A=⎢ ⎢ − k / m1 ⎢ ⎢⎣ k / m2

0 0 k / m1 − k / m2 C = ⎡⎣0

0⎤ ⎥ 1⎥ ; 0⎥ ⎥ 0 ⎥⎦

1 0 0 0

1

⎡ 0 ⎤ ⎥ ⎢ 0 ⎥ B=⎢ ⎢1 / m1 ⎥ ⎥ ⎢ ⎢⎣ 0 ⎥⎦

0 ⎤⎦

0

(4.9.12)

(4.9.13)

Let estimates of parameters k, m1, and m2 be kˆ , mˆ 1, and mˆ 2 , respectively. Then, ⎡ 0 ⎢ 0 ˆ A = ⎢⎢ ˆ − k / mˆ 1 ⎢ ˆ ⎣⎢ k / mˆ 2

0 0

1 0

kˆ / mˆ 1 − kˆ / mˆ

0

2

0⎤ ⎥ 1⎥ ; 0⎥ ⎥ 0 ⎥⎦

0

⎡ 0 ⎤ ⎥ ⎢ 0 ⎥ Bˆ = ⎢ ⎢1 / mˆ 1 ⎥ ⎥ ⎢ ⎢⎣ 0 ⎥⎦

(4.9.14)

Let kˆ = 1, mˆ 1 = 1, and mˆ 2 = 1. When there is no modeling errors, eigenvalues of Asysp are controller poles ( −0.1 ± 1.414 j , 0.1, and 0.1) and observer poles (–0.4 ± 1.414j, –0.4, and –0.4). Associated vectors K and L are as follows: LT = [0.256

1.6

−0.6144

0.96 ]

(4.9.15)

0.004]

(4.9.16)

and K = [0.06

−0.0399

0.4

Let k = 0.5 , kˆ = 1 , mˆ 1 = m1 = 1, and mˆ 2 = m2 = 1. The observer and controller gain vectors L and K are again chosen by Equation 4.9.15 and Equation 4.9.16, respectively. In this case, kˆ ≠ k , and the eigenvalues of Asysp are 1.0085, –0.0881, –0.1231, –1.1567, −0.4989 ± 1.4805 j , and −0.3214 ± 0.0955 j . That is, the closed-loop system has become unstable.

4.9.2 UNMODELED DYNAMICS Let the dynamics of a system be given as x = Ax + Bu

(4.9.17)



225

x R = AR x R + BR u

(4.9.18)

y (t ) = Cx + C R x R

(4.9.19)

and

Here, the controller will be designed on the basis of state equations (4.9.17) only, and state equations (4.9.18) will become residual or unmodeled dynamics. The observer-based controller is u = −K xˆ

(4.9.20)

xˆ = Aˆ xˆ + Bˆ u + L ( y − Cˆxˆ )

(4.9.21)

Because of nonzero BR , the input designed to control the system (4.9.17) will also influence unmodeled dynamics. The term BR is also known as the control spillover. Similarly, because of nonzero C R , the measurement data (Equation 4.9.19) will contain unmodeled dynamics also. The term C R is also known as the observation spillover (Balas, 1978). Define estimated state error vector x as x (t ) = x (t ) − xˆ (t )

(4.9.22)

The complete closed-loop system dynamics is represented by ⎡ x ⎤ ⎡x⎤ ⎢ ⎥ ⎢ ⎥ ⎢ x ⎥ = Asys ⎢ x ⎥ ⎢ x R ⎥ ⎢xR ⎥ ⎣ ⎦ ⎣ ⎦

(4.9.23)

where

Asys

⎡ A − BK ⎢ =⎢ 0 ⎢ − BR K ⎣

BK A − LC BR K

0 ⎤ ⎥ LC R ⎥ AR ⎥⎦

(4.9.24)

When C R = 0 , eigenvalues of Asys are same as those of ( A − BK ) , ( A − LC ), and AR . However, when C R ≠ 0 , eigenvalues of Asys are not same as those of ( A − BK ) , ( A − LC ), and AR , and it is possible to get an unstable closed-loop system in spite of ( A − BK ) , ( A − LC ) , and AR being stable.


226


EXAMPLE 4.11: FLEXIBLE MODE ROBUST CONTROL PROBLEM

AS THE

UNMODELED DYNAMICS

IN THE

BENCHMARK

Assume that m1 = m2 = m . Without an input, Equation 4.9.7 and Equation 4.9.8 can be written in matrix form as + K q = 0 Mq

(4.9.25)

where ⎡m M=⎢ ⎣0

0⎤ ⎥ m⎦

and

⎡k K=⎢ ⎣− k

−k ⎤ ⎥ k ⎦

(4.9.26a,b)

Natural frequencies and modal vectors are computed by solving the following eigenvalue problem: (−ω 2 M + K )q 0 = 0

(4.9.27)

Natural frequencies are found to be ω12 = 0

and

ω 22 =

2k m

(4.9.28)

and modal vectors are as follows: Rigid mode: ⎡⎣1

1⎤⎦

T

corresponding to ω12 = 0

(4.9.29)

corresponding to ω 22 = 2 k / m

(4.9.30)

and Flexible mode: ⎡⎣1

−1⎤⎦

T

Now, a displacement vector can be represented as a linear combination of these modal vectors: ⎡ q1 ⎤ ⎡ r1 ⎤ ⎢ ⎥ = Φ⎢ ⎥ ⎣q2 ⎦ ⎣r2 ⎦

(4.9.31)



227

where ⎡1 Φ=⎢ ⎣1

1⎤ ⎥ −1⎦

(4.9.32)

Substituting (4.9.31) into (4.9.25) with the input and premultiplying both sides by ΦT , r1 =

1 u 2m

(4.9.33)

and

r2 + ω 22 r2 =

1 u 2m

(4.9.34)

Equation 4.9.33 represents the rigid body mode, whereas Equation 4.9.34 represents the vibratory mode. The output equation from (4.9.11) is y (t ) = q2 = r1 − r2

(4.9.35)

Let the controller be designed on the basis of the rigid mode only. Then, the state space model for the controlled mode will be ⎡ p1 ⎤ ⎡0 ⎢ ⎥=⎢ ⎣ p 2 ⎦ ⎣0

1 ⎤ ⎡ p1 ⎤ ⎡ 0 ⎤ ⎥⎢ ⎥+ ⎢ ⎥u 0 ⎦ ⎣ p2 ⎦ ⎣1 / 2 m ⎦

(4.9.36)

where p1 = r1

and

p2 = r1

(4.9.37)

The state space model for the residual mode (4.9.34) will be ⎡ z1 ⎤ ⎡ 0 ⎢ ⎥=⎢ 2 ⎣ z2 ⎦ ⎣ −ω 2

1 ⎤ ⎡ z1 ⎤ ⎡ 0 ⎤ ⎥⎢ ⎥+ ⎢ ⎥u 0 ⎦ ⎣ z2 ⎦ ⎣1 / 2 m ⎦

(4.9.38)

z2 = r2

(4.9.39)

where z1 = r2

and


228


Equation 4.9.35, Equation 4.9.36, and Equation 4.9.38 can be cast as Equation 4.9.17 to Equation 4.9.19 where ⎡0 A=⎢ ⎣0

1⎤ ⎥ 0⎦

⎡ 0 AR = ⎢ 2 ⎣ −ω 2 C = ⎡⎣1

⎡ 0 ⎤ B=⎢ ⎥ ⎣1 / 2 m ⎦

1⎤ ⎥ 0⎦ 0 ⎤⎦

⎡ 0 ⎤ BR = ⎢ ⎥ ⎣1 / 2 m ⎦ C R = − ⎡⎣1

0 ⎤⎦


Consider a simple electromagnetic suspension system shown in Figure P4.1. The electromagnetic force fm is given by I

h0

x (t )

fm

mg

FIGURE P4.1 An electromagnetic suspension system.

fm = α

I2 h2

(P4.1.1)

where I and h are coil current and air gap, respectively. The constant α = μ 0 NAp 2 where μ 0 , N, and Ap are air permeability, number of coil turns, and the face area per single pole of the magnet, respectively. Let h0 be the desired air gap. Then, the current I 0 is calculated from the following static equilibrium condition:

α

I 02 = mg h02

(P4.1.2)



229

Let x(t) be the dynamic displacement of the mass with respect to the static equilibrium position. The nonlinear differential equation of motion is mx = −mg + α

( I 0 + i)2 ( h0 − x )2

(P4.1.3)

where i is the incremental current. Linearizing (P4.1.3) around the static equilibrium condition, mx = k x x + ki i(t )

(P4.1.4)

where kx = 2α

P4.2

I 02 h03

and

ki = 2 α

I0 h02

(P4.1.5a,b)

The system parameters for simulation (Fujita et al., 1995) are as follows: μ 0 = 0.000001258 N / A2 , Ap = 0.000146 m 2, h0 = 0.000508m , m = 0.3 kg. a. Assuming that the output is x(t), design a full-order observer. Present simulation results. b. Assuming that the output is x(t), design a reduced-order observer. Present simulation results. c. Develop a combined observer-controller system with the full-order observer. Your answer should contain a combined observer-controller transfer function. Present simulation results. d. Develop a combined observer-controller system with the reduced-order observer. Your answer should contain a combined observer-controller transfer function. Present simulation results. Consider a plant with the following state space realization: dx = Ax (t ) + bu (t ) ; dt

y (t ) = cx (t )

where ⎡0 A=⎢ ⎣ −3

⎡0 ⎤ 1⎤ ⎥ , b = ⎢ ⎥ , and c = [1 1] 4⎦ ⎣1⎦

a. Develop a combined observer-controller system. Place the state feedback controller poles at –1 and –2, and let both the observer poles be at –4. Give all the equations necessary for implementing your controller in real time.


230

P4.3


b. Find the observer-controller transfer function relating u(s) and y(s). c. Demonstrate the performance of the controller via numerical simulations. Consider the model for longitudinal pressure oscillation in a uniform chamber (Problem P3.19). The outputs yi (t ) (unsteady pressures) are obtained by p sensors located at zsi : n

yi (t ) = p

∑ ψ (z )η (t) ;

si

i = 1, 2, …, p

=1

P4.4

a. Design a full-order observer with n = 4 and p = 1. Assume that the sensor is located at α / 7.5 . b. Assuming that sensors and actuators are collocated at α / 7.5 , design a combined observer-controller system with n = 4 and p = 1. Demonstrate the performance of your controller via numerical simulations. Consider the system dx ⎡1 =⎢ dt ⎣0

1⎤ ⎡0⎤ ⎥ x + ⎢ ⎥ u (t ) + ξ(t ) 1⎦ ⎣1⎦

y (t ) = [1

0 ]x (t ) + θ(t )

where ξ(t ) and θ(t ) are zero mean Gaussian white noise processes with ⎡1 E[ ξ(t )ξT (t + τ )] = σ ⎢ ⎣0

0⎤ ⎥ δ( τ ) 1⎦

E [θ(t )θ(t + τ)] = 1δ(τ)

P4.5

a. Find the steady state Kalman filter gain matrix for two different values of σ , 2 and 0.2. b. Demonstrate the performance of the Kalman filter via numerical simulations for both the values of σ . Discuss your results. Hint: Use Matlab routines: randn and ode23 or ode45. Consider the system dx ⎡1 =⎢ dt ⎣0

1⎤ ⎡0⎤ ⎥ x + ⎢ ⎥ u (t ) + ξ(t ) 1⎦ ⎣1⎦

y (t ) = ⎡⎣1

0 ⎤⎦ x + θ(t )



231

where ξ(t ) and θ(t)) are zero mean Gaussian white noise processes with ⎡0.5 E[ ξ(t )ξT (t + τ )] = ⎢ ⎣0

0⎤ ⎥ δ( τ ) 0.5⎦

E [θ(t )θ(t + τ)] = δ(τ) a. Design the LQG controller to minimize the following objective function: lim E ( y 2 (t ) + u 2 (t )) t →∞

P4.6

b. Demonstrate the performance of your controller via numerical simulations. Hint: Use Matlab routines: ODE45 and randn. Consider the flexible tetrahedral truss structure (Appendix F). The controller is to be designed on the basis of the first four modes of vibration. Design an LQG controller with collocated sensors and actuators such that ⎧⎪ 1 J = lim E ⎨ t f →∞ t ⎪⎩ f

P4.7

tf

∫ 0

⎫⎪ [ yT (t )y(t ) + 0.1uT (t )u(t )dt ]⎬ ⎪⎭

is minimized. Introduce fictitious noises w(t) and θ(t ) (Equation 4.6.1 and Equation 4.6.4), where W = I 8 and Θ = I 3. a. Find the controller transfer matrix and determine its stability. b. Demonstrate the performance of your controller via numerical simulations. For the control of thermoacoustic instability (Annaswamy et al., 2000), a microphone and a loudspeaker are used as a sensor (Figure P4.7) and the actuator, respectively. The plant transfer function is derived to be Loudspeaker Microphone y (s )

u(s ) G (s )

G (s )

K (s )

FIGURE P4.7 Control of thermoacoustic instability.


232


G ( s ) = 2.15 × 10 5

s 2 − 83s + 7.70 × 10 6 ( s + 440 ) ( s + 14.9 )( s 2 + 407 s + 1.03x10 6 ) s 2 − 63s + 9.41 × 10 6

The transfer function of the loudspeaker is given by

G ( s ) =

35.5 s 2 s + 364 s + 3.320 × 10 6 2

a. Design an LQG controller such that lim E ( y 2 (t ) + u 2 (t )) t →∞

P4.8

is minimized. Introduce fictitious noises w(t) and θ(t ) (Equation 4.6.1 and Equation 4.6.4), where W = I and Θ = 1 . b. Demonstrate the performance of your controller via numerical simulations. Consider the single-link manipulator (Figure P4.8a) modeled by Cannon and Schmitz (1984). This model is the same as that in P3.4 except that the mass-moment of inertia of the hub, I h , is included here. Defining IT = I h + Iα , Tip

w( L, t ) torque u(t) Ih

w( x, t ) x (t )

L

FIGURE P4.8A A single-link flexible manipulator.

y ( L, s ) L 1 = + 2 u (s ) IT IT s

∞

∑s i=1

2

βi + 2 ξ i ω i s + ω i2

where I T = 0.44 kg-m2 and L = 1 m. Parameters βi , ξ i , and ω i are experimentally identified and listed in Table P4.8.



233

TABLE P4.8 Parameter Values Mode i

ωi , Hz

ξi

βi

1 2 3

1.8 3.5 8.4

0.05 0.02 0.02

–2.8830 1.4950 –0.5940

a. On the basis of the rigid and the first mode, design an LQG controller klqg (s ) (Banavar and Dominic, 1995) to minimize the following objective function: E (0.25 y 2 ( L, t ) + 3000 u 2 (t )) Introduce fictitious noises w(t) and θ(t ) (Equation 4.6.1 and Equation 4.6.4), where W = 0.1I4 and Θ = 100. b. Find the feedforward function k f in Figure P4.8b such that the unit step response has a zero steady state error. v (t )

kf

k lqg (s )

y ( L, t )

u(t ) G (s )

FIGURE P4.8B Linear quadratic Gaussian (LQG) control system.



5

Robust Control: Fundamental Concepts and H2, H∞, and μ Techniques

First, fundamental concepts for robust control are developed. Then, the robustness of liner quadratic (LQ) and liner quadratic Gaussian (LQG) control techniques developed in Chapter 3 and Chapter 4 are examined. Lastly, theories for H 2, H ∞, and μ techniques are presented along with Bode’s sensitivity integrals and illustrative examples.

5.1 IMPORTANT ASPECTS OF SINGULAR VALUE ANALYSIS 5.1.1 SIGNIFICANCE OF MINIMUM AND MAXIMUM SINGULAR VALUES OF THE TRANSFER FUNCTION MATRIX A FREQUENCY

AT

Consider a multiinput/multioutput (MIMO) system with the p × m transfer function matrix G(s); i.e., y (s ) = G (s )u(s )

(5.1.1)

where y(s ) and u(s ) are output and input vectors, respectively. Let the input vector be sinusoidal with the frequency ω; i.e., u(t ) = e jωt a u

(5.1.2)

where a u = ⎡⎣ au1

au 2

.

.

aum ⎤⎦

T

(5.1.3)

235


236


and the corresponding steady state output vector be y (t ) = e jωt a y

(5.1.4)

where a y = ⎡⎣ ay 1

ay 2

.

ayp ⎤⎦

.

T

(5.1.5)

It is well known that a y = G ( jω )a u

(5.1.6)

Bode magnitude plots for a single input/single output (SISO) system is generalized to a MIMO system by plotting all the singular values (Appendix C) of the matrix G ( jω ) as a function of the frequency ω. The plots of maximum ( σ (G ( jω )) ) and minimum ( σ (G ( jω )) ) singular values have a special significance because of the following relationship (Maciejowski, 1989):

σ (G ( jω )) ≤

G ( jω )a u au

=

a y (ω ) au

≤ σ (G ( jω ))

(5.1.7)

Therefore, the maximum and the minimum singular values of G ( jω ) are upper and lower bounds of the ratio of 2-norms of amplitude vectors of the steady state output and the sinusoidal input of frequency ω. EXAMPLE 5.1.1: SPRING-MASS-DAMPER SYSTEM Consider the spring-mass-damper system shown in Figure 5.1.1. Let the outputs of the system be displacements x1 and x2 of both masses, and the inputs be the forces u1 and u2 acting on masses. The state space model of the system is x = Ax + Bu

(5.1.8)

x1 (t )

x 2 (t )

u1 (t )

u2 ( t )

k1

k1 m

m k2

FIGURE 5.1.1 A spring-mass-damper system.


Robust Control

237

y = Cx + Du

(5.1.9)

where y = ⎡⎣ x1

T

x2 ⎤⎦ ;

⎡ 0 A=⎢ −1 ⎣− M K

x = ⎡⎣ x1

⎤ ⎥; −M Cd ⎦

⎡k + k K=⎢ 1 2 ⎣ − k2

x1

x2

⎡0 ⎤ B=⎢ ⎥; ⎣I ⎦

I

−1

− k2 ⎤ ⎥; k1 + k2 ⎦

x2 ⎤⎦ ;

T

u = ⎡⎣u1

C = ⎡⎣ I

0 ⎤⎦ ;

⎡α Cd = ⎢ ⎣0

0⎤ ⎥; α⎦

⎡m M=⎢ ⎣0

u2 ⎤⎦

T

(5.1.10)

D=0

(5.1.11)

0⎤ ⎥ m⎦

(5.1.12)

The transfer function matrix relating inputs and outputs is described as follows: y (s ) = G (s )u(s )

(5.1.13)

G (s ) = C (sI − A)−1 B + D

(5.1.14)

where

is a square matrix of order 2. At any frequency ω, there will be two singular values of G ( jω ) . These singular values provide bounds of the magnitude of the frequency response (Equation 5.1.7). Using the Matlab command “sigma,” singular values are plotted in Figure 5.1.2 as a function of the frequency ω for m = 1 kg, k1 = 100 N / m , k2 = 500 N / m , and α = 1 N − sec/ m . It is interesting to observe that peaks of singular values are located near the natural frequencies of the system.

5.1.2 AN INEQUALITY

FOR

ROBUSTNESS TEST

Let A be a nonsingular matrix; i.e., σ( A) > 0

(5.1.15)

where σ( A) is the minimum singular value of the matrix A. Let us try to find a perturbation matrix L of the smallest size in the sense of a 2-norm such that the matrix (A+L) is singular or rank deficient. In this case, there exists a nonzero vector x such that x =1 2

and

( A + L )x = 0

(5.1.16a,b)


238


0

Max. σ(G(jω))

Singular Values (dB)

–20

–40

–60 Min. σ(G(jω))

–80

–100

–120 100

101

102

103

Frequency (rad/sec)

FIGURE 5.1.2 Singular values of the mechanical system in Figure 5.1.1.

Therefore, using properties of singular values (Appendix C), σ ( A) ≤ Ax = Lx ≤ L 2

2

2

= σ(L )

(5.1.17)

where σ( L ) is the maximum singular value of the matrix L. In other words, the singularity of the matrix (A+L) implies σ ( A) ≤ σ ( L ); i.e., the size of the perturbation matrix must be at least σ( A) . This analysis can also be viewed as the following condition for the nonsingularity of the matrix (A+L): σ ( A) > σ ( L )

(5.1.18)

The condition (5.1.18) serves as an inequality for the development of various robustness criteria.

5.2 ROBUSTNESS: SENSITIVITY AND COMPLEMENTARY SENSITIVITY 5.2.1 BASIC DEFINITIONS In Figure 5.2.1, the transfer function matrix for a MIMO system is denoted by G (s ) . The controller transfer function matrix is denoted by K (s ). Other variables are described as follows:


Robust Control

239

d(s ) r (s )

u(s )

e (s )

K (s )

y (s ) G (s )

y ( s ) d( s )

n(s )

FIGURE 5.2.1 A multiinput/multioutput system with disturbances.

n(s ) : d(s ) : r(s ) : e(s ) :

Measurement or sensor noise vector Disturbance vector at the plant output Reference (external) input vector Tracking error vector

Analyzing the Figure 5.2.1 (Anderson and Moore, 1990), it is clear that y (s ) − d(s ) = G (s ) K (s )e(s )

(5.2.1)

e(s ) = r(s ) − ( y (s ) + n(s ))

(5.2.2)

and

Substituting (5.2.2) into (5.2.1), y (s ) = To (s )[r(s ) − n(s )] + So (s )d(s )

(5.2.3)

To (s ) = ( I + G (s ) K (s ))−1 G (s ) K (s )

(5.2.4)

So (s ) = ( I + G (s ) K (s ))−1

(5.2.5)

e(s ) = r(s ) − n(s ) − d(s ) − ( y (s ) − d(s ))

(5.2.6)

e(s ) = So (s )(r(s ) − d(s )) − So (s )n(s )

(5.2.7)

where

and

In addition, from (5.2.2),

Using (5.2.1),


240


Also, u(s ) = K (s )e(s ) = K (s ) So (s )(r(s ) − d(s )) − K (s ) So (s )n(s )

(5.2.8)

Definitions 5.2.1 So (s ) = ( I + G (s ) K (s ))−1 is called the output sensitivity function. (5.2.9) To (s ) = ( I + G (s ) K (s ))−1 G (s ) K (s ) is called the output complementary sensitivity function. (5.2.10) Fact 5.2.1 To (s ) = ( I + G (s ) K (s ))−1 G (s ) K (s ) = G (s ) K (s )( I + G (s ) K (s ))−1

(5.2.11)

Proof ( I + GK )−1 GK = [(GK )−1 ( I + GK )]−1 = [(GK )−1 + I ]−1 = [( I + GK )(GK )−1 ]−1 = GK ( I + GK )−1 This completes the proof. Fact 5.2.2 So (s ) + To (s ) = I

(5.2.12)

Proof So (s ) + To (s ) = ( I + GK )−1 + ( I + GK )−1 GK = ( I + GK )−1 ( I + GK ) = I This completes the proof. The relationship (5.2.12) represents a fundamental trade-off in the design of a control system. It implies that both So and To(s) cannot be simultaneously small. Note the following observations (Anderson and Moore, 1990):


Robust Control

241

1. Assuming that d(s ) and n(s ) are zero, Equation 5.2.7 yields e(s ) = So (s )r(s )

where

e(s ) = r(s ) − y (s )

(5.2.13)

The error signal e(s ) = r(s ) − y (s ) is the tracking error. Therefore, the sensitivity function So (s ) should be small for a small tracking error. More specifically, the following condition should be satisfied for a good tracking: σ ( So ( jω )) 1 + ε

(5.2.53)

EXAMPLE 5.2.2: A SPRING-MASS SYSTEM Consider the spring-mass-damper system (Figure 5.1.1) again. First, an lqg controller, R = ρc I and Q = C T C in Equation 4.8.9, is designed (Matlab Program 5.2.1) to obtain the controller transfer function K(s). Then, the sensitivity function So (s ), the complementary sensitivity function To (s ), and K (s ) So (s ) are plotted for two values (0.01 and 0.1) of the control weighting ρc in Figure 5.2.5, Figure 5.2.6, and Figure 5.2.7, respectively. The sensitivity function for ρc = 0.01 is smaller than that for ρc = 0.1 near the natural frequencies of the system. Because of the constraint So (s ) + To (s ) = I , the complementary sensitivity function for ρc = 0.01 is larger than that for ρc = 0.1 near the natural frequencies of system. But at all frequencies, the control input or K (s ) So (s ) for ρc = 0.1 is significantly smaller than that for ρc = 0.01 . MATLAB PROGRAM 5.2.1: LQG CONTROL % clear all close all % k1=100; k2=500;

OF A

SPRING-MASS SYSTEM


248

Linear Systems: Optimal and Robust Control So(s), ρc=0.01

So(s), ρc=0.1

0.5

0.05 0

0

–0.05



–0.5

–1

–0.1

–0.15

–1.5

–0.2

–0.25

–2 –0.3 –2.5

–3 10–1

–0.35

102 100 101 Frequency (rad/sec)

103

–0.4 10–1


103

FIGURE 5.2.5 Output sensitivity functions for linear quadratic Gaussian control of a springmass-damper system.

K=[k1+k2 -k2;-k2 k1+k2]; al=1; % A=[zeros(2,2),eye(2);-K -al*eye(2)]; B=[0 0 1 0;0 0 0 1]'; C=[1 0 0 0;0 1 0 0]; D=zeros(2,2); % sysG=ss(A,B,C,D); % for iplot=1:2 %LQG Controller rhoc=0.001*(10)îplot;


Robust Control

249 K(s)So(s), ρc=0.1

0

–20

–20

–40

–40

–60

–60

–80 Singular Values (dB)


K(s)So(s), ρc=0.01

–80 –100 –120

–100 –120 –140

–140

–160

–160

–180

–180

–200

–200 10–1


103

–220 10–1


103

FIGURE 5.2.6 Output complementary sensitivity functions for linear quadratic Gaussian control of a spring-mass-damper system.

V=[eye(4) zeros(4,2);zeros(2,4) 0.01*eye(2)]; W=[C'*C zeros(4,2);zeros(2,4) rhoc*eye(2)]; [Af,Bf,Cf,Df]=lqg(A,B,C,D,W,V); sysK=ss(Af,Bf,Cf,Df); % sysId=ss([],[],[],eye(2)); % sysGK=series(sysK,sysG); % %Sensitivity Function sys_Sens=feedback(sysId,sysGK); figure(1) subplot(1,2,iplot)


250

Linear Systems: Optimal and Robust Control K(s)So(s), ρc=0.1

20

0

10

–10

0

–20

–10

–30



K(s)So(s), ρc=0.01

–20 –30 –40

–40 –50 –60

–50

–70

–60

–80

–70 10–1


103

–90 10–1


103

FIGURE 5.2.7 Control inputs or for linear quadratic Gaussian control of a spring-massdamper system.

sigma(sys_Sens) [sv_Sens,freq]=sigma(sys_Sens); grid xlabel('Frequency','Fontsize',12) ylabel('Singular Values','Fontsize',12) if(iplot==1) title('S_o(s), \rho_c=0.01','Fontsize',12) end if(iplot==2) title('S_o(s), \rho_c=0.1','Fontsize',12) end % %


Robust Control

251

%Complementary Sensitivity Function sys_CSens=feedback(sysGK,sysId); [A_CSens,B_CSens,C_CSens,D_CSens]=branch(sys_CSens); eig(A_CSens) figure(2) subplot(1,2,iplot) sigma(sys_CSens,freq) grid xlabel('Frequency','Fontsize',12) ylabel('Singular Values','Fontsize',12) if(iplot==1) title('T_o(s), \rho_c=0.01','Fontsize',12) end if(iplot==2) title('T_o(s), \rho_c=0.1','Fontsize',12) end % figure(3) subplot(1,2,iplot) %K(s)S(s) sys_KSens=series(sys_Sens,sysK); sigma(sys_KSens,freq) grid xlabel('Frequency','Fontsize',12) ylabel('Singular Values','Fontsize',12) if(iplot==1) title('K(s)S_o(s), \rho_c=0.01','Fontsize',12) end if(iplot==2) title('K(s)S_o(s), \rho_c=0.1','Fontsize',12) end end


252


0

1 X

x(s )

B

(s)

K

u (s )

FIGURE 5.3.1 A linear quadratic regular feedback loop.

5.3. ROBUSTNESS OF LQR AND KALMAN FILTER (KF) FEEDBACK LOOPS 5.3.1 LINER QUADRATIC REGULATOR (LQR) FEEDBACK LOOP Consider the state space model x = Ax + Bu(t )

(5.3.1)

u(t ) = − Kx (t )

(5.3.2)

and the state feedback law

where K is the optimal state feedback gain matrix (Chapter 3, Section 3.5). Taking Laplace transforms of (5.3.1) and (5.3.2) with zero initial conditions, u(s ) = − Kx (s )

(5.3.3)

x (s ) = Φ(s ) Bu(s )

(5.3.4)

Φ(s ) = (sI − A)−1

(5.3.5)

Relationships (5.3.3) and (5.3.4) can be represented as an LQR feedback loop shown in Figure 5.3.1. The LQR loop transfer matrix GLQ (s ) is given by GLQ (s ) = K Φ(s ) B

(5.3.6)

Breaking the loop at (1) in Figure 5.3.1, Return difference matrix = I + GLQ (s )

(5.3.7)

Fact 5.3.1: Return Difference Equality [ I + GLQ (−s )]T R[ I + GLQ ( s )] = R + [ N Φ(−s )B]T [ N Φ( s )B]

(5.3.8)


Robust Control

253

where Q = NTN

(5.3.9)

Proof Consider the algebraic Riccati equation (ARE) (Equation 3.5.55 with N = 0): − PA − AT P + PBR −1BT P = Q

(5.3.10)

Equation 5.3.10 can be written as P (sI − A) + (−sI − AT ) P + K T RK = Q

(5.3.11)

K = R −1BT P

(5.3.12)

P Φ −1 (s ) + [ΦT (−s )]−1 P + K T RK = Q

(5.3.13)

because

Using (5.3.5),

Left-multiplying (5.3.13) by BT ΦT (−s ), and right-multiplying (5.3.13) by Φ(s ) B , BT ΦT (−s ) PB + BT P Φ(s ) B + BT ΦT (−s ) K T RK Φ(s ) B = BT ΦT (−s )Q Φ(s ) B (5.3.14) From (5.3.12), RK = BT P

(5.3.15)

Substituting (5.3.15) into (5.3.14), BT ΦT (−s )K T R + RK Φ( s )B + BT ΦT (−s )K T RK Φ( s )B

(5.3.16)

= BT ΦT (−s )QΦ( s )B Adding R to both sides of Equation 5.3.16, using Equation 5.3.6 and Equation 5.3.9, T T R + GLQ (−s ) R + RGLQ (s ) + GLQ (−s ) RGLQ (s ) = R + BT ΦT (−s ) NN T Φ(s ) B (5.3.17)


254


Equation 5.3.17 can be written as T [ I + GLQ (−s )]R[ I + GLQ (s )] = R + [ N Φ(−s ) B]T [ N Φ(s ) B]

(5.3.18)

This completes the proof.

5.3.2 GAIN

AND

PHASE MARGINS

OF A

SINGLE-INPUT SYSTEM

For a single-input system, GLQ (s ) and R are scalars. Let R =ρ>0

(5.3.19)

1 T [1 + GLQ (−s )][1 + GLQ (s )] = 1 + [ N Φ(−s ) B]T [ N Φ(s ) B] ρ

(5.3.20)

Equation 5.3.18 reduces to

Substituting s = jω into (5.3.20), 2

1 + GLQ ( jω ) = 1 +

2 1 N Φ( jω ) B ρ

(5.3.21)

Therefore, 1 + GLQ ( jω ) ≥ 1

(5.3.22)

The Nyquist map of GLQ (s ) will be outside the circle of unit radius centered at −1 + j 0 (Figure 5.3.2). Let 1 be the factor by which the magnitude of the input can change. In this case, the loop transfer function will be 1GLQ (s ) . Because the Nyquist map of GLQ (s ) will be outside the circle of unit radius centered at −1 + j 0 , it will only intersect the negative real axis beyond the point G (–2, 0). Hence, the number of encirclements of −1 + j 0 will remain unchanged provided 1 >

1 2

(5.3.23)

Therefore, the gain margin of the single-input LQR feedback loop shown in Figure 5.3.1 is at least between ∞ and 0.5. Let φ1 be the angle by which the phase of the input can change. In this case, the loop transfer function will be e jφ1 GLQ (s ) . Because the Nyquist map of GLQ (s )


Robust Control

255

1 2

j

Im ag

3 2

P1

600

G 2

1

600

1

Re al

P2

1 2

j

3 2

FIGURE 5.3.2 Forbidden unit circle with center at (–1,0).

will be outside the circle of unit radius centered at −1 + j 0 and unit circles centered ⎛ 1 3⎞ at (0,0) and (–1,0) intersect (Figure 5.3.2) each other at P1 ⎜⎜− , ⎟⎟ and ⎝ 2 2 ⎠ ⎛ 1 3⎞ P2 ⎜⎜− , − ⎟ , the number of encirclements of (–1, 0) will remain unchanged 2 ⎟⎠ ⎝ 2 provided φ1 ≤ 60°

(5.3.24)

Therefore, the phase margin of the single-input LQR feedback loop shown in Figure 5.3.1 is at least 60°. EXAMPLE 5.3.1: LQ CONTROL

OF A

SIMPLE MASS

Consider the double integrator in Example 4.8. With ρc = 0.05, GLQ (s ) =

2.9907s + 4.4721 s2

(5.3.25)

Therefore, GLQ ( jω ) = −

4.4721 2.9907 −j ω ω2

(5.3.26)


256


Imag.

R 1

0 s-plane

6 2

5

Real 4

R

3

FIGURE 5.3.3A Nyquist path.

and ∠GLQ ( jω ) = tan −1

−2.9907ω −4.4721

(5.3.27)

Setting GLQ ( jω ) = 1, ω 4 − (2.9907)2 ω 2 − ( 4.4721)2 = 0

(5.3.28)

ω = ω p = 3.2858 rad/sec

(5.3.29)

Solving (5.3.28),

and ∠GLQ ( jω p ) = tan −1

−2.9907ω p −4.4721

= 245.53°

(5.3.30)

Therefore, Phase Margin = 245.53 – 180 = 65.53°

(5.3.31)

From the Nyquist map (Figures 5.3.3a and 5.3.3b), it is clear that the gain margin is between 0 and ∞.


Robust Control

257

Imag. G LQ (s ) plane

4 5 Real center : (-1,0) radius = 1

1,2,3 6

FIGURE 5.3.3B Nyquist map of GLQ (s ) .

5.3.3 GAIN

AND

PHASE MARGINS

OF A

MULTIINPUT SYSTEM

Assume that R = ρI

(5.3.32)

1 [ I + GLQ (−s )]T [ I + GLQ (s )] = I + [ N Φ(−s ) B]T [ N Φ(s ) B] ρ

(5.3.33)

Then, from (5.3.8),

Substituting s = jω into (5.3.33), 1 [ I + GLQ (− jω )]T [ I + GLQ ( jω )] = I + [ N Φ(− jω ) B]T [ N Φ(( jω ) B] (5.3.34) ρ Therefore, [ I + GLQ (− jω )]T [ I + GLQ ( jω )] ≥ I

(5.3.35)

Fact 5.3.2 Consider the system shown in Figure 5.3.4. If W H + W > I , I + GLQ ( jω )W ( jω ) is nonsingular (Anderson and Moore, 1990) where W H is the complex conjugate transpose of W.


258


W (s )

G LQ (s )

FIGURE 5.3.4 Linear quadratic loop with an uncertainty W(s).

Proof: (by contradiction) Let I + GLQ ( jω )W ( jω ) be singular. In this case, there will be a nontrivial vector q such that ( I + GLQ ( jω )W ( jω ))q = 0

(5.3.36)

GLQ Wq = −q

(5.3.37)

H H GLQ + GLQ + GLQ GLQ ≥ 0

(5.3.38)

or

From (5.3.35),

H where GLQ ( jω ) is the complex conjugate transpose of GLQ ( jω ) . Premultiplying and postmultiplying (5.3.38) by q H W H and Wq , respectively,

H H q H W H GLQ Wq + q H W H GLQ Wq + q H W H GLQ GLQ Wq ≥ 0

(5.3.39)

Using (5.3.37), −q H Wq − q H W H q + q H q ≥ 0

(5.3.40)

q H (W + W H − I )q ≤ 0

(5.3.41)

or

Relation (5.3.41) contradicts the assumption W H + W > I in the statement of Fact 5.3.2. It should be noted that the Nyquist condition for the stability of the system in Figure 5.3.4 is that I + GLQ ( jω )W ( jω ) is nonsingular. Therefore, the system shown in Figure 5.3.4 is stable when W H + W > I .


Robust Control

259

ᐉ1 ᐉ2 . +

.

GLg(s)

.

–

ᐉm

FIGURE 5.3.5 A multiinput/multioutput linear quadratic regular loop with a magnitude uncertainty in each input channel.

Special Case Let W be a diagonal matrix, with diagonal elements being 1 , 2, …, and m . Then, the condition W H + W > I reduces to Hi + i > 1 ;

i = 1, 2,…, m

(5.3.42)

Conditions (5.3.42) yield guaranteed levels of gain and phase margins of the LQR feedback loop as follows. Gain Margin Assume that 1 , 2 , …, and m are real numbers. Then, conditions (5.3.42) yield

i >

1 ; 2

i = 1, 2,…, m

(5.3.43)

The LQR feedback loop (Figure 5.3.5) has a guaranteed infinite upward gain margin, and 1/2 downward gain margin in each input channel independently and simultaneously (Athans, 1992). Phase Margin Assume that i = e jφi Then, conditions (5.3.42) yield

(5.3.44)


260


e jφ1

e jφ2 .

+

. –

GLg(s)

. e jφn

FIGURE 5.3.6 A multiinput/multioutput linear quadratic regular loop with a phase uncertainty in each input channel.

e− jφi + e jφi = 2 cos φ i > 1

(5.3.45)

φ i < 60°

(5.3.46)

Therefore,

The LQR feedback loop (Figure 5.3.6) has a guaranteed ±60° phase margin in each input channel independently and simultaneously (Athans, 1992). Note that the stability is not guaranteed if both a phase shift within ±60° and a ⎛1 ⎞ gain change in the interval ⎜ ,∞ ⎟ are simultaneously introduced in an input channel. ⎝2 ⎠

5.3.2 KALMAN FILTER (KF) LOOP A Kalman filter (Chapter 4, Section 4.6) is represented by xˆ = Axˆ (t ) + Bu(t ) + L ( y − yˆ )

(5.3.47)

yˆ (t ) = Cxˆ (t )

(5.3.48)

where

Taking the Laplace transform of (5.3.47) with zero initial conditions, xˆ (s ) = Φ(s ) Bu (s ) + Φ(s ) L ( y (s ) − yˆ (s )) where Φ(s ) = (sI − A)−1, Equation (5.3.5).

(5.3.49)


Robust Control

261

u(t ) B

xˆ (t )

y (t ) L

(s )

yˆ (t )

C

FIGURE 5.3.7 A Kalman filter feedback loop.

The relations (5.3.48) and (5.3.49) are depicted in Figure 5.3.7. The Kalman filter loop transfer matrix GKF (s ) is given as GKF (s ) = C Φ(s ) L

(5.3.50)

Similar to Equation 5.3.8, it can be proved that ( I + GKF (s ))Θ( I + GKF (−s ))T = Θ + (C Φ(s ) N f )(C Φ(−s ) N f )T

(5.3.51)

W = N f N Tf

(5.3.52)

where

Assume that Θ = χI ;

χ>0

(5.3.53)

1 (C Φ(s ) N f )(C Φ(−s ) N f )T χ

(5.3.54)

From (5.3.51) to (5.3.53), ( I + GKF (s ))( I + GKF (−s ))T = I + Substituting s = jω in (5.3.54), ( I + GKF ( jω ))( I + GKF (− jω ))T = I +

1 (C Φ( jω ) N f )(C Φ(− jω) N f )T (5.3.55) χ

As the second term on the right-hand side of (5.3.55) is positive semidefinite, ( I + GKF ( jω )( I + GKF (− jω ))T ≥ I

(5.3.56)


262


or σ ( I + GKF ( jω )) ≥ 1

(5.3.57)

Because the sensitivity function of the Kalman filter (KF) loop is S KF = ( I + GKF )−1

(5.3.58)

Equation 5.3.57 yields σ ( S KF ( jω )) ≤ 1 ;

∀ω

(5.3.59)

This property indicates that the KF loop will never amplify output disturbances at any frequency. The complementary sensitivity function, TKF , is given by TKF = I − S KF

(5.3.60)

From (5.3.59) and (5.3.60) and properties of singular values (Appendix C), σ (TKF ( jω )) ≤ 2 ;

∀ω

(5.3.61)

Using (5.2.45), the stability of the KF loop is guaranteed to multiplicative modeling uncertainty provided σ ( Δ( jω )) < 0.5

(5.3.62)

where Δ(s ) relates the actual plant transfer function GΔ (s ) and the model transfer matrix as follows: GΔ (s ) = ( I + Δ(s ))G (s ) EXAMPLE 5.3.2: KALMAN FILTER

FOR A

(5.3.63)

SIMPLE MASS

Consider the double integrator in Example 4.6. With ρ = 0.05, GKF (s ) =

4.4721s + 10 s2

(5.3.64)


Robust Control

263

0

SKF

–10

Magnitude (dB)

–20

–30

–40

–50

–60

–70 10–1

100

101

102

Frequency (rad/sec)

FIGURE 5.3.8 Sensitivity function of the Kalman filter loop (Example 5.3.2).

Therefore,

S KF =

s2 s 2 + 4.4721s + 10

and

TKF =

4.4721s + 10 (5.3.65a,b) s 2 + 4.4721s + 10

Both sensitivity and complementary sensitivity functions satisfy (5.3.59) and (5.3.61), respectively (Figure 5.3.8 and Figure 5.3.9).

5.4 LQG/LTR CONTROL 5.4.1 LACK

OF

GUARANTEED ROBUSTNESS

OF

LQG CONTROL

Consider the following state space realization (Doyle, 1978): ⎡1 A=⎢ ⎣0

1⎤ ⎥; 1⎦

Let the weighing matrices be

⎡0 ⎤ B=⎢ ⎥ ⎣1⎦

and

C = ⎡⎣1

0 ⎤⎦

(5.4.1)


264


5

0 TKF

Magnitude (dB)

–5

–10

–15

–20

–25

–30 10–1

100

101

102

Frequency (rad/sec)

FIGURE 5.3.9 Complementary sensitivity function of the Kalman filter loop (Example 5.3.2).

⎡1⎤ Q = ρ ⎢ ⎥ ⎡⎣1 ⎣1⎦

1⎤⎦ ;

R =1

(5.4.2)

⎡1⎤ W = σ ⎢ ⎥ ⎡⎣1 ⎣1⎦

1⎤⎦ ;

Θ=1

(5.4.3)

and

Solving the control algebraic Riccati equation (CARE), Equation 4.8.19, K = α ⎡⎣1

1⎤⎦

where

α = 2+ 4+ρ

(5.4.4)

Solving the filter algebraic Riccati equation (FARE), Equation 4.8.16, L = β ⎡⎣1

1⎤⎦

where

β = 2+ 4+σ

Let the actual plant dynamics be represented by

(5.4.5)


Robust Control

265

x = Ax (t ) + χBu (t )

(5.4.6)

where χ ≠ 1 represents the uncertainty in the plant model. The system matrix for the closed-loop system can be written as 0 ⎤ ⎥ − χα ⎥ 1 ⎥ ⎥ 1 − α ⎥⎦

(5.4.7)

det(sI − Asys ) = s 4 + p1s 3 + p2 s 2 + p3s + (1 + (1 − χ)αβ) = 0

(5.4.8)

Asys

⎡ A =⎢ ⎣ LC

⎡1 ⎢ ⎤ − χBK ⎢0 = ⎥ A − BK − LC ⎦ ⎢β ⎢ ⎢⎣β

1 1 0 0

0 − χα 1− β −α − β

The characteristic equation of Asys can be written as

where coefficients p1, p2 , and p3 can be determined. The closed-loop system is unstable if 1 + (1 − χ)αβ < 0

(5.4.9)

The condition (5.4.9) leads to χ > 1+

1 αβ

(5.4.10)

With large values of α and β, even a slight increase in the value of χ from its nominal value will render the closed-loop system to be unstable. In other words, the gain margin of the LQG control system can be almost zero. This example clearly shows that the robustness of the LQG control system to modeling errors is not guaranteed.

5.4.2 LOOP TRANSFER RECOVERY (LTR) LTR stands for loop transfer recovery. The LQ control or KF loop has an excellent robustness property. But the LQG control system (Figure 5.4.1) does not have any guaranteed robustness. Therefore, the LQG/LTR control tries to recover a target loop transfer function GTL (s ), which is chosen as the KF loop here: GTL (s ) = GKF (s )

(5.4.11)

It should be noted that the noise vectors in the Kalman filter (Chapter 4, Section 4.6) can be considered “fictitious,” and its statistics W and Θ are treated as design parameters to achieve the desired properties of the target loop.


266


r (s )

K lqg (s )

u(s )

G (s )

y (s )

FIGURE 5.4.1 A multiinput/multioutput feedback loop with linear quadratic Gaussian controller.

The LQG controller transfer function matrix is given by K lqg (s ) = − K (sI − A + BK + LC )−1 L

(5.4.12)

where K and L are obtained from (4.8.18) and (4.8.15), respectively. From (4.8.18), with R = ρI and Q = C T C , K = ρ−1BT P

(5.4.13)

PA + AT P − ρ−1PBBT P + C T C = 0

(5.4.14)

where

Theorem 5.4.1 Assume that G (s ) = C (sI − A)−1 B is a square matrix and has only minimum phase transmission zeros. If K is obtained by (5.4.13) and L by (4.8.15), pointwise in s, lim C ( sI − A)−1 BK ( sI − A + BK + LC )−1 L = C ( sI − A)−1 L ρ→0

(5.4.15)

Equation (5.4.15) states that lim G ( s )K lqg ( s ) = GTL ( s ) ρ→0

(5.4.16)

Proof (Triantafyllou and Hover, 2000) For a square and minimum phase G(s) (Kwakernaak and Sivan, 1972), lim K ρ = HC ρ→0

(5.4.17)

where H is an orthonormal matrix. Using the matrix inversion lemma (Appendix A),


Robust Control

267

((sI − A + LC ) + BK )−1 = Λ − ΛB( I + K ΛB)−1 K Λ

(5.4.18)

Λ = (sI − A + LC )−1

(5.4.19)

where

Therefore, K lqg (s ) = K ((sI − A + LC ) + BK )−1 L = [ I − K ΛB( I + K ΛB)−1 ]K ΛL (5.4.20) Using a special case of the matrix inversion lemma (Appendix A), K lqg (s ) = ( I + K ΛB)−1 K ΛL = ( ρI + ρK ΛB)−1 ρK ΛL

(5.4.21)

Using (5.4.17), lim K lqg ( s ) = ( HC ΛB)−1 HC ΛL = (C ΛB)−1 H −1HC ΛL = (C ΛB)−1 C ΛL

(5.4.22)

C Λ = C (sI − A + LC )−1 = [ I − C Φ(s ) L ( I + C Φ(s ) L )−1 ]C Φ(s )

(5.4.23)

Φ(s ) = (sI − A)−1

(5.4.24)

ρ→0

Next,

where

Again, using a special case of the matrix inversion lemma (Appendix A), C Λ = ( I + C Φ(s ) L )−1C Φ(s )

(5.4.25)

Substituting (5.4.25) into (5.4.22), lim K lqg ( s ) = [( I + C Φ( s )L )−1 G ( s )]−1 ( I + C Φ( s )L )−1 C Φ( s )L

(5.4.26)

G (s ) = C Φ(s ) B

(5.4.27)

ρ→0

where


268


Simplifying (5.4.26), lim K lqg ( s ) = G −1 ( s )C Φ( s )L

(5.4.28)

lim G ( s )K lqg ( s ) = C ( sI − A)−1 L

(5.4.29)

ρ→0

From (5.4.28),

ρ→0

This proves the theorem. In other words, the loop transfer function in the Figure 5.4.1 will converge to the target loop transfer matrix. The LQG/LTR design method was introduced (Doyle and Stein, 1981; Athans, 1986) before the development of the H ∞ method, which is a more general approach and can directly handle many types of modeling uncertainties. EXAMPLE 5.4.1: A SIMPLE MASS

OR A

DOUBLE INTEGRATOR

Consider the double integrator plant: ⎡0 A=⎢ ⎣0

1⎤ ⎥; 0⎦

⎡0 ⎤ B=⎢ ⎥; ⎣1⎦

and

C = [1

0]

(5.4.30)

Let K = [ k1

k2 ]

and

LT = [ 1

2 ]

(5.4.31)

Then, K (sI − ( A − BK − LC ))−1 L =

( k1 1 + k2 2 )s + ( k1 2 + k2 22 − k2 1 2 ) s 2 + ( k2 + 1 )s + ( k2 1 + k1 + 2 )

(5.4.32)

From (5.4.13), k1 = ρ−1/ 2

and

k2 = 2 ρ−1/ 4

(5.4.33)

Substituting (5.4.33) into (5.4.32) and taking limit, lim K ( sI − ( A − BK − LC ))−1 L = 1s + 2 ρ→0

Therefore,

(5.4.34)


Robust Control

269

lim C ( sI − A)−1 BK ( sI − A + BK + LC )−1 L = ρ→0

1s + 2 s2

(5.4.35)

Note that C (sI − A)−1 L =

1s + 2 s2

(5.4.36)

Hence, Equation 5.4.16 is verified, or the loop transfer function has been recovered.

5.5 H2 AND H∞ NORMS 5.5.1 DEFINITION

OF

L2 NORM

Consider a vector signal x(t), which is a function of time t ≥ 0 . Then, the L2 norm of this vector signal x(t) is defined as ∞

2

x

L2

=

∫ x (t)x(t)dt < ∞ T

(5.5.1)

0

5.5.2 DEFINITION

OF

H2 NORM

Consider a transfer function matrix G(s) with each element being a strictly proper and stable transfer function. The H 2 norm of the transfer function matrix is given by +∞

G (s )

2 H2

∫

1 Trace G ( jω )G H ( jω ) d ω = 2π

(5.5.2)

−∞

Using Parseval’s theorem (Appendix B), ∞

G (s )

2 H2

∫

= Trace G (t )G T (t ) dt

(5.5.3)

0

where G(t) is the inverse Laplace transform of G(s). Using the definition of singular values (Appendix C), Equation 5.5.2 can also be written as

G (s )

2 H2

1 = 2π

+∞

p

∫ ∑ [σ (G ( jω))] dω 2

i

−∞ i =1

(5.5.4)


270


where σ i (G ( jω )) is the ith singular value of G ( jω ) , and p is the number of nonzero singular values. It is also clear from the definition of singular value, σ i (G ( jω )) = σ i (G (− jω ))

(5.5.5)

Furthermore, for the integral in (5.5.4) to be finite, σ i (G ( jω )) → 0 as ω → ∞ for each i. This condition is satisfied by a strictly proper transfer function matrix.

5.5.3 SIGNIFICANCE I.

OF

H2 NORM

Connection with Unit Impulse Responses

Consider the MIMO system shown in Figure 5.5.1, for which y (s ) = G (s )u(s )

(5.5.6)

Let a unit impulse function, δ(t ), be applied in one input channel at a time. If this input is applied in the ith channel, the corresponding output vector can be written as y i (s ) = G (s )v i

(5.5.7)

where v i = [0

.

.

0

1

0

.

0 ]T

(5.5.8)

Now, ∞

y i (t )

2 L2

=

∫ Trace[y (t)y i

T i

(5.5.9)

(t )]dt

0

Using Parseval’s theorem (Appendix B),

y i (t )

u1 u2

2 L2

1 = 2π

∞

∫ Trace[y ( jω)y i

H i

( jω )]d ω

−∞

y1 . .

.

G (s )

.

um

FIGURE 5.5.1 A multiinput/multioutput linear system.

y2

. .

yp

(5.5.10)


Robust Control

271

Using (5.5.7),

y i (t )

2 L2

1 = 2π

∞

∫ Trace[G ( jω)v v G T i i

H

( jω )]d ω

(5.5.11)

−∞

Using (5.5.8),

2

y i (t ) L = 2

1 2π

∞

∫ [ G ( j ω) i1

2

2

+ … + Gip ( jω) ]dω

(5.5.12)

−∞

Summing up over all input channels, m

∑

2

y i (t ) L = 2

i =1

1 2π

⎡ ⎛ Trace ⎢G ( jω) ⎜ ⎜ ⎢ ⎝ ⎣ −∞ ∞

m

∑

∫

i =1

⎤ ⎞ vi viT ⎟ G H ( jω)⎥ dω ⎟ ⎥ ⎠ ⎦

(5.5.13)

Using (5.5.8), It can be shown that m

∑v v

T i i

= Im

(5.5.14)

i =1

Therefore, the equation (5.5.13) can be written as m

∑ i=1

y i (t )

2 L2

=

1 2π

∞

∫

Trace[G ( jω )G H ( jω )]d ω = G (s )

2 H2

(5.5.15)

−∞

Therefore, minimizing the H 2 norm of the transfer function matrix is equivalent to minimizing the sum of squares of L2 norm of outputs due to unit impulse input in each channel. II.

Connection with the Root Mean Square (RMS) Response

In time domain, t

y (t ) =

∫ G (t − τ)u(τ)d τ 0

(5.5.16)


272


Let each input be an independent zero-mean white noise with unit intensity; i.e., E (u(t )u T (τ)) = I m δ(t − τ)

(5.5.17)

The mean square response can be written (Stein, 1988) as E ( y T (t ) y (t )) = E (Trace[ y (t ) y T (t )])

(5.5.18)

or ⎛ ⎡ E ( yT (t )y(t )) = E ⎜Trace ⎢ ⎜ ⎢⎣ ⎝ ⎛ E ( yT (t )y(t )) = Trace ⎜ ⎜ ⎝

∫

t

G (t − τ )u( τ )d τ

0

t

t

0

0

∫ ∫

⎛ E ( y (t )y(t )) = Trace ⎜ ⎜ ⎝ T

t

∫ 0

⎤⎞ uT (λ )G T (t − λ )d λ⎥⎟ (5.5.19) ⎥⎦⎟ ⎠

⎞ G (t − τ )E (u( τ )uT (λ ))G T (t − λ )d τd λ ⎟ (5.5.20) ⎟ ⎠

t

t

0

0

∫ ∫

⎞ G (t − τ )G T (t − λ )δ( τ − λ )d τd λ ⎟ ⎟ ⎠

⎛ E ( yT (t )y(t )) = Trace ⎜ ⎜ ⎝

t

∫ 0

⎞ G (t − λ )G T (t − λ )d λ ⎟ ⎟ ⎠

(5.5.21)

(5.5.22)

Introducing ν=t−λ

(5.5.23)

t

E ( y T (t ) y (t )) =

∫ Trace[G (ν)G

T

( ν)]d ν

(5.5.24)

0

∞

lim E ( yT (t )y(t )) = t →∞

∫ Trace[G(ν)G (ν)]d ν T

0

Using Parseval’s theorem,

(5.5.25)


Robust Control

273

lim E ( yT (t )y(t )) = t →∞

1 2π

∞

∫ Trace[G( jω)G

H

2

( jω)]dω = G ( s ) H

2

(5.5.26)

−∞

Therefore, minimizing the H 2 norm of the transfer function matrix is equivalent to minimizing the RMS of outputs in the statistical steady state due to independent zero-mean white noise inputs.

5.5.4 DEFINITION

OF

H∞ NORM

Consider a transfer function matrix G(s) with each element being a proper and stable transfer function. The H ∞ norm of the transfer function matrix is given by G (s )

H∞

=

sup ω

σ (G ( jω ))

(5.5.27)

where supremum (sup) is defined in Appendix C. Bounded Real Lemma The condition G (s )

H∞

ρi

(6.4.7)

otherwise

where ρi is the boundary layer thickness around the ith hyperplane, illustrated in Figure 6.2. The system is linear inside the boundary layers. From (6.4.5) through (6.4.7), ⎡ ⎛ ⎤ η ⎞ u(t ) = −(GB)−1 ⎢G ⎜ A − LC + I n ⎟ xˆ (t ) + GLy(t )⎥ μ ⎠ ⎣ ⎝ ⎦

(6.4.8)

Without any loss of generality, ηi and ρi have been assumed to be identical for each hyperplane and have been set equal to η and ρ, respectively. Furthermore, inside boundary layers, si = −νsi

(6.4.9)

where ν=

η ρ

(6.4.10)

For state estimation, xˆ (0 ) is usually set to zero. In this case, s(0 ) = 0 , and the solution of (6.4.9) indicates that s(t ) = 0 . In other words, the system is always on the intersection of all sliding hyperplanes, and the resulting control input is always linear as described by (6.4.8). An important point to note here is that the control input depends on estimated states and directly on outputs as well (Sinha and Millar, 1995). This is a significant difference with the standard observer based control system where inputs only depend on estimated states. From (6.4.3) and (6.4.8), the controller transfer function is as follows: u(s ) = − K es (s ) y (s )

(6.4.11)

K es (s ) = Ζ 1 (sI n − A + BΖ 1 + LC )−1 Ζ 2 + (GB)−1 GL

(6.4.12)

Ζ 1 = (GB)−1 G ( A + νI n − LC )

(6.4.13)

where


396


Ζ 2 = ( L − ΩL )

(6.4.14)

Ω = B(GB)−1 G

(6.4.15)

Using (6.4.1) through (6.4.3), and (6.4.8), the dynamics of the closed-loop system is described as ⎡ x ⎤ ⎡ A − Ω( A + νI n ) ⎢⎥ = ⎢ 0 ⎣ x ⎦ ⎣

Ω( A + νI n ) − ΩLC ⎤ ⎡ x ⎤ ⎥⎢ ⎥ A − LC ⎦ ⎣ x ⎦

(6.4.16)

where x = x − xˆ

(6.4.17)

Equation 6.4.16 shows that the eigenvalues of the closed-loop system are those of A − Ω( A + νI n ) and A − LC . Therefore, the sliding hyperplane matrix G and the boundary layer parameter ν , which constitute the controller parameters, can be chosen independently of the observer gain matrix L. In other words, the separation property holds, and the controller and observer can be designed independently. Fact The m eigenvalues of A − Ω( A + νI n ) are −ν and the remaining n−m eigenvalues can be arbitrarily placed in the complex plane by a proper choice of G (Sinha and Miller, 1995) as the system (A,B) is controllable.

6.5 OPTIMAL SLIDING MODE GAUSSIAN (OSG) CONTROL State Dynamics x (t ) = Ax (t ) + Bu(t ) + w(t )

(6.5.1)


(6.5.2)

E ( w(t ) w T (t + τ)) = W δ(τ)

(6.5.3)

and


Robust Control: Sliding Mode Methods

397

The matrix W is called the intensity matrix with the property W = W T ≥ 0. Measurement Equation y(t ) = Cx(t ) + θ(t )

(6.5.4)

where y(t ) is the sensor noise vector, and θ(t ) is a continuous-time Gaussian white noise vector. It is assumed that E (θ(t )) = 0

(6.5.5)

E (θ(t )θT (t + τ )) = Θδ( τ )

(6.5.6)

and

where Θ = ΘT > 0. It is also assumed that process and measurement noise vectors are uncorrelated; i.e., E (θ(t )wT (t + τ )) = E (θ(t ))E (wT (t + τ )) = 0

(6.5.7)

Theorem (Sinha and Miller, 1995) For the stochastic system (6.5.1)–(6.5.7), the matrix G chosen on the basis of Equation 6.2.37 minimizes the following objective function: ⎧⎪ 1 J = lim E ⎨ t f →∞ t ⎪⎩ f

tf

∫ 0

⎫⎪ xT (t )Qx(t )dt ⎬ ⎪⎭

(6.5.8)

when states are estimated by the Kalman filter (Chapter 4, Section 4.6); i.e., xˆ = Axˆ + Bu + L ( y − Cxˆ )

(6.5.9)

L = ΣC T Θ −1

(6.5.10)

AΣ + ΣAT + W − ΣC T Θ −1C Σ = 0

(6.5.11)

and

It should be noted here that the constraints for the minimization of the objective function (6.5.8) are


398


s(t ) = Gxˆ (t ) = 0

(6.5.12)

J = lim E {xˆ T (t )Qxˆ (t )} + tr ( ΣQ )

(6.5.13)

Proof Using (4.8.11),

t →∞

As tr ( ΣQ ) is independent of the input u(t), the minimization of Equation 6.5.13 is equivalent to minimizing the following objective function: J m = lim E {xˆ T (t )Qxˆ (t )} t →∞

(6.5.14)

Define the following similarity transformation: qˆ (t ) = Hxˆ (t )

(6.5.15)

where the matrix H is same as (6.2.12). From (6.5.9) and (6.5.15), qˆ = Aqˆ + Bu + Lv (t )

(6.5.16)

where A and B are the same as defined by (6.2.14) and (6.2.15); and L = HL

(6.5.17)

v (t ) = y (t ) − Cxˆ (t )

(6.5.18)

Similar to Equation 6.2.16, the vector qˆ is again partitioned as ⎡ qˆ ⎤ qˆ = ⎢ 1 ⎥ ⎣qˆ 2 ⎦

(6.5.19)

The first (n−m) equations of (6.5.16) can be written as qˆ 1 = A11qˆ 1 + A12 qˆ 2 + Lnm v (t )

(6.5.20)

where the matrix Lnm is composed of the first (n−m) rows of the matrix L . Similar to (6.2.35), J m = lim E {qˆ 1T (t )Qr qˆ 1 (t ) + 2 qˆ 1T Nqˆ 2 + qˆ T2 Rqˆ 2 } t →∞

(6.5.21)



399

As v(t ) in Equation 6.5.20 is a white noise vector, the stochastic regulator theory described in Chapter 4, Section 4.7 is directly applicable here. Therefore, the following relationship will yield the optimal solution: qˆ 2 (t ) = − Kqˆ 1 (t )

(6.5.22)

where K is given by (6.2.37). Lastly, the matrix G is given by Equation 6.2.19: G = ⎡⎣ K

I m ⎤⎦ H

(6.5.23)

This completes the proof. EXAMPLE 6.4: A THREE-DISK SYSTEM (ECP MANUAL) Consider the three-disk system shown in Figure 5.11.4. Define the state vector as θ (t )⎤⎦

xT (t ) = ⎡⎣θ(t )

(6.5.24)

The state equations are derived to be x = Ax (t ) + bτ(t )

(6.5.25)

where ⎡ 0 A=⎢ −1 ⎣− M C

I3 ⎤ ⎥; − M −1K ⎦

⎡ 0 ⎤ b = ⎢ −1 ⎥ ⎢⎣ M B f ⎥⎦

(6.5.26a,b)

Let the output be y (t ) = θ1 (t ). In this case, y (t ) = Cx (t )

(6.5.27)

where C = ⎡⎣1

0

0

0

0

0 ⎤⎦

(6.5.28)

Take W = I 6 and Θ = 1. In this case, the Kalman filter gain vector is K F = ⎡⎣0.9732

0.6177

0.7304

−0.0264

−0.4088

0.3570 ⎤⎦

T

(6.5.29)


400


To design the sliding mode control, Q = C TC

(6.5.30)

From the decomposition (6.2.34), Qr , N, and R are computed. Because R turns out to be zero, it is arbitrarily chosen to be 10–7. The OSG controller transfer function, Equation 6.4.12, is plotted in Figure 6.4. The variance of the output, E ( y 2 ) , is found to be 1.0356. Next, an LQG controller is designed by taking Q = C T C and R = 0.01 in Equation 4.8.8. The LQG controller transfer function, Equation 4.3.11, is also plotted in Figure 6.4. The variance of the output, E ( y 2 ) , is found to be 0.9898. Therefore, both LQG and OSG controllers lead to similar levels of performance. This is also confirmed by the singular value plots of the closed-loop transfer matrices T relating the 7 × 1 stochastic disturbance vector [ w T (t ) θ(t )] to the output, y(t), illustrated in Figure 6.5. Observing locations of peaks and troughs in Figure 6.4, controller transfer functions seem to be inversions of the open loop transfer function. However, this inversion phenomenon is less severe around the natural frequency ω 3 = 65.39 rad/sec. in the case of the OSG controller transfer function, in Figure 6.6. Similar results were also found for a flexible spacecraft (Sinha and Miller, 1995). 40


20

Open Loop

0 LQG –20

–40

OSG

–60

–80 100

101

102 Frequency (rad/sec)

FIGURE 6.4 OSG and LQG controller transfer functions.

103



401

40

Open Loop


20

0 LQG –20 OSG –40

–60

–80 100

101

102

103

Frequency (rad/sec)

FIGURE 6.5 OSG and LQG closed-loop transfer functions. 0

–5

–10 Singular Values (dB)

Open Loop –15

–20

–25

LQG

–30

–35

–40 1017

OSG

1018 Frequency (rad/sec)

1019

FIGURE 6.6 OSG and LQG controller transfer functions in a small frequency range.


402



Consider the single-link flexible manipulator, Figure P3.4, with a single vibratory mode. Define the state vector as q0

x = ⎡⎣q0

q1

q1 ⎤⎦

T

Design a sliding hyperplane matrix to locate the eigenvalues at −σ , − σ ± jω1 . Choose σ such that the damping ratio in the vibratory mode equals 0.1. P6.2 Consider the single-link flexible manipulator, Figure P3.4, with a single vibratory mode. Define the state vector as q0

x = ⎡⎣q0

q1

q1 ⎤⎦

T

Design a sliding hyperplane matrix to minimize the following objective function: ∞

∫ x xdt T

0

P6.3

Consider the single-link flexible manipulator, Figure P3.4, with a single vibratory mode. The output is y(L,t). Define the state vector as x = ⎡⎣q0

P6.4

q0

q1

q1 ⎤⎦

T

Taking W = I and Θ = 0.5 , design an OSG controller such that E ( x T x ) is minimized. Demonstrate the performance of the controller via numerical simulations. Design an LQG control system with a comparable value of E ( y 2 (t )) . Compare the OSG and LQG sensitivity and controller transfer functions. Consider the Problem 4.8. On the basis of the rigid and the first mode, design an OSG controller kosg (s ) to minimize E ( y 2 ) . If the weighting of equivalent input is zero, assume it to be a small number. Introduce fictitious noises w(t) and θ(t ) , Equation 4.6.1 and Equation 4.6.4, where W = 0.1I 4 and Θ = 100 . Demonstrate the performance of the controller via numerical simulations. Design an LQG control system with a comparable value of E ( y 2 (t )) . Compare the OSG and LQG sensitivity and controller transfer function.



P6.5

403

Consider the flexible tetrahedral truss structure (Appendix F). The controller is to be designed on the basis of the first four modes of vibration. Introduce fictitious noises w(t) and θ(t ) , Equation 4.6.1 and Equation 4.6.4, where W = I 8 and Θ = I 3 . a. Design an OSG controller with collocated sensors and actuators, such that J = E (y T y) is minimized. If the weighting of equivalent input is zero, assume it to be a small number. b. Design an LQG controller with collocated sensors and actuators such that J = E ( y T y + ρu T u)

P6.6

is minimized. Select ρ such that E ( y T y ) is almost equal to that for OSG control. c. Compare the controller transfer matrices in part (a) and part (b). d. Compare the closed-loop transfer matrices in part (a) and part (b). Consider the OSG and LQG controllers designed in P6.5. Assume that parameters are uncertain according to (6.3.34) and (6.3.35). a. Find the maximum range of each parametric uncertainty for robust stability using the small gain theorem. b. Find the maximum range of each parametric uncertainty for robust stability using μ-analysis.



References Abate, M., Barmish, B.R., Murillo-Sanchez, C., and Tempo, R., Applications of some new tools to robust stability analysis of spark ignition engines: a case study, IEEE Transactions on Control Systems Technology, Vol. 2, No. 1, 22–30, 1994. Annaswamy, A.M., Fleifil, M., Rumsey, J.W., Prasanth, R., Hathout, J.-P., and Ghoneim, A.F., Thermoacoustic instability: model-based optimal control designs and experimental validation, IEEE Transactions on Control Systems Technology, Vol. 8, No. 6, 905–918, 2000. Anderson, B.D.O. and Moore, J.B., Optimal Control: Linear Quadratic Methods, PrenticeHall, Englewood Cliffs, NJ, 1990. Asada, H. and Slotine, J.J.E., Robot Analysis and Control, John Wiley & Sons, NY, 1986. Astrom, K.J., Klein, R.E., and Lennartsson, A., Bicycle Dynamics and Control, IEEE Control Systems Magazine, Vol. 25, No. 4, 26–47, 2005. Athans, M., A tutorial on the LQG/LTR method, Proceedings of American Control Conference, 1289–1296, 1986. Athans, M., Notes for Multivariable Control Systems Course, MIT, Cambridge, MA, 1992. Balas, M.J., Feedback control of flexible systems, IEEE Transactions on Automatic Control, Vol. AC-23, No. 4, 673–679, 1978. Banavar, R. and Dominic, P., An LQG/H∞ controller for a flexible manipulator, IEEE Transactions on Control Systems Technology, Vol. 3, No. 4, 409–416, 1995. Biegler, L., Class Notes of Advanced Process Control Course, Carnegie Mellon University, Pittsburgh, PA, 1982. Braatz, R. and Morari, M., Robust control for a noncollocated spring-mass system, AIAA Journal of Guidance, Control and Dynamics, Vol. 15, No. 5, 1103–1110, 1992. Bryson, A.E., Some connections between modern and classical control concepts, ASME Journal of Dynamic Systems, Measurement and Control, Vol. 101, 91–98, June 1979. Cannon, R.H. and Schmitz, E., Initial experiments on the end-point control of a flexible onelink robot, International Journal of Robotic Research, Vol. 3, No. 3, 1984. Chen, C-T., Systems and Signal Analysis, Saunders College Publishing: Holt, Reinhart and Winston, NY, 1989. Chiang, R.Y. and Safonov, M.G., H∞ synthesis using a bilinear pole shifting transform, AIAA Journal of Guidance, Control and Dynamics, Vol. 15, No. 5, 1111–1117, 1992a. Chiang, R.Y. and Safonov, M.G., Robust Control Toolbox — User’s Guide, The MathWorks, 1992b (www.mathworks.com). Dahleh, M., Notes on μ-analysis (Multivariable Control Systems Course), MIT, Cambridge, MA, 1992. Dorato, P., Abdallah, C., and Cerone, V., Linear Quadratic Control: An Introduction, PrenticeHall, Englewood Cliffs, NJ, 1995. Doyle, J.C., Guaranteed margins for LQG regulators, IEEE Transactions on Automatic Control, Vol. 23, No. 4, 756–757, August 1978. Doyle, J.C. and Stein, G., Multivariable feedback design: concepts for a classical/modern synthesis, IEEE Transactions on Automatic Control, Vol. AC-26, No. 1, February 1981. 405


406


Doyle, J.C., Lenz, K., and Packard, A., Design examples using μ-synthesis: space shuttle lateral axis FCS during reentry, Proceedings of NATO Workshop: Modelling, Robustness and Sensitivity Reduction in Control Systems, Curtain, R.F., Ed., 1986, pp. 127–154. Doyle, J.C., Glover, K., Khargonekar, P.P., and Francis, B.A., State-space solutions to standard H2 and H∞ control problems, IEEE Transactions on Automatic Control, Vol. AC-34, No. 8, 831–847, 1989. Drazenovic, B., The invariance condition in variable structure systems, Automatica, Vol. 5, 287–295, 1969. ECP, Educational Control Products (ECP) Manual, http://www.ecpsystems.com/. Friedland, B., Control System Design — An Introduction to State Space Analysis, McGrawHill, NY, 1985. Fujita, M., Namerikawa, T., Matsumura, F., and Uchida, K., μ-synthesis of an electromagnetic suspension system, IEEE Transaction on Automatic Control, Vol. 40, No. 3, 530–536, 1995. Gopal, M., Modern Control System Theory, John Wiley & Sons, NY, 1984. Green, M. and Limebeer, D.J.N., Linear Robust Control, Prentice-Hall, Englewood Cliffs, NJ, 1995. Gupta, N.K., Frequency-shaped cost functionals: extension of linear-quadratic-Gaussian design methods, AIAA Journal of Guidance, Control and Dynamics, Vol. 3, No. 6, 529–535, November–December 1980. Hughes, R.O., Optimal control of sun tracking solar concentrators, ASME Journal of Dynamic Systems, Measurement and Control, Vol. 101, 157–161, June 1979. Kailath, T., Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980. Kalman, R.E. and Bucy, R.S., New results in linear filtering and prediction theory, ASME Journal of Basic Engineering, Vol. 83, 95–107 1961. Kao, C.K., Robust Control of Vibration in Flexible Structures, Ph.D. dissertation, The Pennsylvania State University, University Park, PA, 1989. Kao, C.K. and Sinha, A., 1991, Sliding mode control of vibration in flexible structures using estimated states, Proceedings of American Control Conference, Vol. 3, 1991, pp. 2467–2474. Kar, I.N., Miyakura, T., and Seto, K., Bending and torsional vibration control of a flexible plate structure using H∞-based robust control law, IEEE Transactions on Control Systems Technology, Vol. 8, No. 3, 545–553, 2000. Kitada, H., Kondo, O., Kusachi, H., and Sasame, K., H∞ control of molten steel level in continuous caster, IEEE Transactions on Control Systems Technology, Vol. 6, No. 2, 200–207, 1998. Kuo, B.C., Automatic Control Systems, Prentice-Hall, Englewood Cliffs, NJ, 1995. Kwakernaak, H. and Sivan, R., Linear Optimal Control Systems, John Wiley & Sons, 1972. Lowell, J. and McKell, H.D., The stability of bicycles, American Journal of Physics, Vol. 50, No. 12, 1982, 1106–1112, 1982. Lublin, L. and Athans, M., Linear quadratic regulator control, The Control Handbook, 1996, pp. 635–650. Maciejowski, J.M., Multivariable Feedback Design, Addison-Wesley, Workingham, England, 1989. Meditch, J.S., Stochastic Optimal Linear Estimation and Control, McGraw-Hill, New York, 1969. Pai, M.C. and Sinha, A., Generating command inputs to eliminate residual vibration via direct optimization and quadratic programming techniques, Active control of vibration and noise, DE-Vol. 93, Proceedings of the 1996 ASME International Mechanical Engineering Congress and Exposition, Atlanta, GA, November 1996.


References

407

Pai, M.C. and Sinha, A., Active control of vibration in a flexible structure with uncertain parameters via sliding mode and H ∞ / μ techniques, Vibration and noise control, DEVol. 97, DSC-Vol. 65, Proceedings of the 1998 ASME International Mechanical Engineering Congress and Exposition, Anaheim, CA, November 1998. Pai, M.C. and Sinha, A., Sliding mode control of vibration in a flexible structure via estimated states and H ∞ / μ techniques, Proceedings of American Control Conference, Chicago, IL, June 2000, pp. 1118–1123. Pao, L. and Singhose, W.E., Robust minimum time control of flexible structures, Automatica, Vol. 34, No. 2, 229–236, 1998. Papoulis, A., Probability, Random Variables, and Stochastic Processes, McGraw-Hill, NY, 1965. Ray, W.H., Advanced Process Control, McGraw-Hill, 1981. Reddy, B.D., Functional Analysis and Boundary-Value Problems: An Introductory Treatment, Longman Scientific and Technical, U.K., 1986. Rugh, W., Linear System Theory, Prentice-Hall, Englewood Cliffs, NJ, 1993. Safanov, M.G., Imaginary-axis zeros in multivariable H∞-optimal control, Proceedings of NATO Workshop: Modelling, Robustness and Sensitivity Reduction in Control Systems, Curtain, R.F., Ed., 1986, pp. 71–81. Sage, A.P. and White, C.C., Optimum Systems Control, Prentice-Hall, Englewood Cliffs, NJ, 1977. Schultz, M.J. and Inman, D.J., Eigenstructure assignment and controller optimization for mechanical systems, IEEE Transactions on Control Systems Technology, Vol. 2, No. 2, 88–100, 1994. Seo, B., and Chen, C.T., The relationship between the Laplace transform and the Fourier transform, IEEE Transactions on Automatic Control, Vol. AC-31, No. 8, 751, August 1986. Singer, N.C. and Seering, W.P., Preshaping command inputs to reduce system vibration, ASME Journal of Dynamic Systems, Measurement and Control, Vol. 112, 76–82, 1990. Singhose, W. and Pao, L., A comparison of input shaping and time-optimal flexible body control, Control Engineering Practice, Vol. 5, No. 4, 459–467, 1997. Sinha, A. and Miller, D.W., Optimal sliding mode control of a flexible spacecraft under stochastic disturbances, AIAA Journal of Guidance, Dynamics and Control, Vol. 118, No. 3, 486–492, 1995. Stein, G., Formal Control System Synthesis with H 2 / H ∞ Criteria, Lecture Notes for Multivariable Control Systems Course, MIT, Cambridge, MA, 1988. Stein, G., Respect the Unstable, IEEE Control Systems Magazine, Vol. 23, No. 4, 12–25, 2003. Strang, G., Linear Algebra and Its Applications, Harcourt Brace Jovanovich Publishers, San Diego, CA, 1988. Strunce, R., Lin, J., Hegg, D., and Henderson, T., Actively Controlled Structures Theory, C.S. Draper Lab., Cambridge, MA, Final Report Vol. 2, R-1338, 1979. Thompson, A.G., An active suspension with optimal linear state feedback, Vehicle System Dynamics, Vol. 5, 187–203, 1976. Triantafyllou, M.S. and Hover, F.S., Loop transfer recovery, Maneuvering and Control of Marine Vehicles, MIT, Cambridge, MA, 2000 (www.mit.edu and www.core.org.cn). Tsach, U., Koontz, J.W., Ignatoski, M.A., and Geselowitz, D.B., An adaptive aortic pressure observer for the Penn State electric ventricular assist device, IEEE Transactions on Biomedical Engineering, Vol. 37, No. 4, 374–383, April 1990. Utkin, V.I., Variable structure systems with sliding modes: a survey, IEEE Transactions on Automatic Control, Vol. AC-22, 212–222, 1977.


408


Utkin, V.I. and Yang, K.D., Methods for constructing discontinuity planes in multidimensional variable structure systems, Automation and Remote Control, Vol. 31, 1466–1470, 1978. Vidyasagar, M., Control System Synthesis: A Factorization Approach, MIT Press, Cambridge, MA, 1985. Vidyasagar, M., Nonlinear Systems Analysis, 2nd ed., Prentice-Hall, Englewood Cliffs, NJ, 1993. Wiberg, D.M., State space and linear systems, Schaum’s Outline Series, McGraw-Hill, New York, 1971. Wie, B. and Bernstein, D.S., Benchmark problems for robust control design, AIAA Journal of Guidance, Control and Dynamics, Vol. 15, No. 5, 1057–1059, 1992. Yang, V., Sinha, A., and Fung, Y.T., State feedback control of longitudinal combustion instabilities, AIAA Journal of Propulsion and Power, Vol. 8, No. 1, 66–73, 1992. Zhou, K. and Doyle, J.C., Essentials of Robust Control, Prentice-Hall, Upper Saddle River, NJ, 1998. Zhou, K., Doyle, J.C., and Glover, K., Robust and Optimal Control, Prentice-Hall, Upper Saddle River, NJ, 1996.


Appendix A: Linear Algebraic Equations, Eigenvalues/Eigenvectors, and Matrix Inversion Lemma A.1 SYSTEM OF LINEAR ALGEBRAIC EQUATIONS (GOPAL, 1984; STRANG, 1988) Consider the solution of the following system of linear algebraic equations: Ax = b

(A.1)

where A is an m × n matrix. Given an m-dimensional vector b, the objective is to find the n-dimensional vector x. It should be noted that m and n need not be equal.

DEFINITIONS Linear Independence of Vectors The vectors x1 , x 2 ,…, x n are linearly independent provided the condition α1x1 + α 2 x 2 + … + α n x n = 0 is true only when scalars α1 = α 2 = … = α n = 0 . Range or Column Space Range or column space consists of all combinations of column vectors of a matrix. Rank of a Matrix The rank of a matrix is the maximum number of linearly independent columns, which is also the dimension of the range or column space of A. The maximum number of linearly independent columns always equals the maximum number of linearly independent rows of a matrix. Therefore, the rank of a matrix is also the maximum number of linearly independent rows. The rank of a matrix also equals the dimension of the largest nonsingular square array.

409


410


Null Space The null space of the matrix A consists of all vectors y satisfying Ay = 0 . The dimension of the null space, also known as nullity, is the maximum number of independent solutions of Ay = 0 . Theorem A.1 Let A be an m × n matrix. Then, γ = n−ρ

(A.2)

where ρ and γ are the rank and nullity of the matrix A. EXAMPLE A.1 ⎡1 ⎢ A = ⎢1 ⎢2 ⎣

3 2 5

5 3 8

7⎤ ⎥ 4⎥ 11⎥⎦

(A.3)

The last row is the sum of first two rows, which are independent. Therefore, rank(A) = 2

(A.4)

From (A.2), the dimension of the null space = 2. There are two independent solutions of Ay = 0, which can be written as follows after ignoring the last row:

⎡1 ⎢ ⎣1

3 2

5 3

⎡ y1 ⎤ ⎢ ⎥ 7 ⎤ ⎢ y2 ⎥ ⎡ 0 ⎤ =⎢ ⎥ ⎥ 4 ⎦ ⎢ y3 ⎥ ⎣ 0 ⎦ ⎢ ⎥ ⎣⎢ y 4 ⎦⎥

(A.5)

or y1 + 3 y2 + 5 y3 + 7 y 4 = 0

(A.6)

y1 + 2 y2 + 3 y3 + 4 y 4 = 0

(A.7)

We have two equations in four unknowns. Two unknowns can be chosen arbitrarily, and then the two remaining unknowns can be determined uniquely. Equations can be expressed as


Appendix A

411

⎡1 ⎢ ⎣1

3 ⎤ ⎡ y1 ⎤ ⎡ −5 y3 − 7 y 4 ⎤ ⎥ ⎥⎢ ⎥ = ⎢ 2 ⎦ ⎣ y2 ⎦ ⎣ −3 y3 − 4 y 4 ⎦

(A.8)

Solving for y1 and y2 , ⎡ y1 ⎤ ⎡1 ⎢ ⎥=⎢ ⎣ y2 ⎦ ⎣1

−1

3⎤ ⎡−5 y3 − 7 y4 ⎤ ⎡ y3 + 2 y4 ⎤ ⎥=⎢ ⎥ ⎥ ⎢ 2⎦ ⎣−3y3 − 4 y4 ⎦ ⎣−2 y3 − 3y4 ⎦

(A.9)

Choosing y3 = 1 and y4 = 0,

y b1

⎡ y1 ⎤ ⎡ 1 ⎤ ⎢ ⎥ ⎢ ⎥ y −2 = ⎢ 2⎥= ⎢ ⎥ ⎢ y3 ⎥ ⎢ 1 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢⎣ y 4 ⎥⎦ ⎢⎣ 0 ⎥⎦

(A.10)

yb2

⎡ y1 ⎤ ⎡ 2 ⎤ ⎢ ⎥ ⎢ ⎥ y −3 = ⎢ 2⎥= ⎢ ⎥ ⎢ y3 ⎥ ⎢ 0 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢⎣ y 4 ⎥⎦ ⎢⎣ 1 ⎥⎦

(A.11)

Choosing y3 = 0 and y 4 = 1,

Solutions y b1 and y b2 are independent of each other. Therefore, any vector in the null space can be written as y = αy b1 + βy b 2

(A.12)

where α and β are arbitrary real numbers. Theorem A.2 The system of equations (A.1) is solvable if and only if the vector b belongs to the range or column space of the matrix A; i.e., rank ([ A b]) = rank ( A)

(A.13)

Theorem A.3 Consider the system of equations (A.1), which satisfies the solvability condition (A.13).


412


a. Irrespective of whether m is greater than or equal to or less than n, the number of independent equations will always be less than or equal to the number of unknowns. In other words, effectively, m ≤ n . b. When rank(A) = n, Equation A.1 has the unique solution x = A −1b

(A.14)

c. When rank ( A) < n , the dimension of the null space is greater than zero (Equation A.2). In this case, there are infinite many solutions of Equation A.1, which can be written as x = x p + xh

(A.15)

where x p is a particular solution satisfying Ax p = b , and x h is any solution in the null space of the matrix A; i.e., Ax h = 0 . Note that there are infinite many nonzero vectors in the null space. As a result, there are infinite many solutions for x. Example: ⎡1 ⎢ A = ⎢1 ⎢2 ⎣

3 2 5

5 3 8

7⎤ ⎥ 4⎥; 11⎥⎦

⎡4⎤ ⎢ ⎥ b = ⎢3 ⎥ ⎢7⎥ ⎣ ⎦

(A.16)

The vector b is the sum of the first two columns of the matrix A. Therefore, b belongs to the column or the range space of the matrix A, and Equation A.1 is solvable. As shown in Example A.1, rank(A) = 2, and ignoring the last row,

⎡1 ⎢ ⎣1

3 2

5 3

⎡ x1 ⎤ ⎢ ⎥ 7 ⎤ ⎢ x2 ⎥ ⎡ 4 ⎤ =⎢ ⎥ ⎥ 4 ⎦ ⎢ x3 ⎥ ⎣ 3 ⎦ ⎢ ⎥ ⎢⎣ x 4 ⎥⎦

(A.17)

Rewriting this equation as ⎡1 ⎢ ⎣1 Solving for x1 and x2 ,

3 ⎤ ⎡ x1 ⎤ ⎡ 4 − 5 x3 − 7 x 4 ⎤ ⎥ ⎥⎢ ⎥ = ⎢ 2 ⎦ ⎣ x2 ⎦ ⎣ 3 − 3 x3 − 4 x 4 ⎦

(A.18)


Appendix A

413

⎡ x1 ⎤ ⎡1 ⎢ ⎥=⎢ ⎣ x2 ⎦ ⎣1

−1

3 ⎤ ⎡ 4 − 5 x3 − 7 x 4 ⎤ ⎡ 1 + x3 + 2 x 4 ⎤ ⎥=⎢ ⎥ ⎥ ⎢ 2 ⎦ ⎣ 3 − 3 x3 − 4 x 4 ⎦ ⎣1 − 2 x3 − 3 x 4 ⎦

(A.19)

A particular solution is generated by arbitrarily setting x3 = x 4 = 0 : ⎡ x1 ⎤ ⎡ 1 ⎤ ⎢ ⎥ ⎢ ⎥ x 1 xp = ⎢ 2 ⎥ = ⎢ ⎥ ⎢ x3 ⎥ ⎢ 0 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢⎣ x 4 ⎥⎦ ⎢⎣0 ⎥⎦

(A.20)

The independent solutions in the null space have already been generated (A.10 and A.11). Therefore, all the infinite many solutions can be expressed as ⎡1⎤ ⎡1⎤ ⎡2⎤ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ 1 −2 −3 x = ⎢ ⎥+α⎢ ⎥+β⎢ ⎥ ⎢0 ⎥ ⎢1⎥ ⎢0⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢⎣0 ⎥⎦ ⎢⎣ 0 ⎥⎦ ⎢⎣ 1 ⎥⎦

(A.21)

where α and β are arbitrary real numbers. Theorem A.4 Consider the system of equations (A.1), which does not satisfy the solvability condition (A.3). In this case, m > n , and the following error vector ε will always be nonzero: ε = Ax − b ≠ 0

(A.22)

x = ( AT A)−1 AT b

(A.23)

Then, the solution

minimizes the following least squares error LSE: m

LSE =

∑ε i=1

2 i

= εT ε

(A.24)


414


A SYSTEMATIC TEST

FOR

LINEAR INDEPENDENCE

VECTORS

OF

Consider a set of vectors x1 , x 2 ,…, x n , with each vector having m elements. First, a scalar or inner product of any two vectors is defined as follows: m

x i , x j = x iT x j = x Tj x i =

∑x x i

j

(A.25)

=1

To test linear independence of vectors, the following relationship is considered: α1x1 + α 2 x 2 + … + α n x n = 0

(A.26)

Taking the scalar product of (A.26) with the vector x i , α1 x i , x1 + α 2 x i , x 2 + … + α n x i , x n = 0;

i = 1, 2,…, n

(A.27)

Putting (A.27) in matrix form, Gα α=0

(A.28)

where ⎡ x 1, x 1 ⎢ ⎢ x 2 , x1 ⎢ G=⎢ . ⎢ . ⎢ ⎢⎣ x n , x1

x 1, x 2

.

.

x2 , x2

.

.

. . x n , x2

. . .

. . .

x 1, x n ⎤ ⎥ x2 , x n ⎥ ⎥ . ⎥ ⎥ . ⎥ x n , x n ⎥⎦

(A.29)

and α = [α1

α2

.

.

αn ]

T

(A.30)

Note that the matrix G is symmetric. From (A.28), α = 0 if and only if det G ≠ 0

(A.31)

Therefore, vectors x1, x 2 ,…, x n are independent if and only if det G ≠ 0 . detG is known as Gramian or Gram determinant.


Appendix A

415

A.2 EIGENVALUES AND EIGENVECTORS (STRANG, 1988) The eigenvalue/eigenvector relationship for a square matrix A of order n is given by Av = λv

(A.32)

where λ is an eigenvalue, and v is the corresponding eigenvector. Equation A.32 can also be written as (λI − A)v = 0

(A.33)

det(λI − A) = 0

(A.34)

For nontrivial solution of v,

Equation A.34 is a polynomial equation of order n. Therefore, there are n roots or n eigenvalues of the matrix A. Let these eigenvalues be denoted as λ1, λ 2 ,…, λ n . Then, the eigenvector v i corresponding to the eigenvalue λ i is determined by solving (λ i I − A)v i = 0

(A.35)

It should be noted that the eigenvector v i belongs to the null space of (λ i I − A) .

A.3 MATRIX INVERSION LEMMA ( A + BCD )−1 = A −1 − A −1B(C −1 + DA −1B)−1 DA −1

(A.36)

Proof Consider the following relationship: ( A + BCD )( A + BCD )−1 = I

(A.37)

A( A + BCD )−1 + BCD ( A + BCD )−1 = I

(A.38)

( A + BCD )−1 = A −1 − A −1BCD ( A + BCD )−1

(A.39)

or

or


416


or ( A + BCD )−1 = A −1 − A −1B[ A( I + A −1BCD )(CD )−1 ]−1

(A.40)

( A + BCD )−1 = A −1 − A −1B[ A(CD )−1 + A −1B)]−1

(A.41)

( A + BCD )−1 = A −1 − A −1B( D −1C −1 + A −1B)−1 D −1DA −1

(A.42)

( A + BCD )−1 = A −1 − A −1B[ D ( D −1C −1 + A −1B)]−1 DA −1

(A.43)

( A + BCD )−1 = A −1 − A −1B(C −1 + DA −1B)−1 DA −1

(A.44)

or

or

or

or

This completes the proof. Special Case Let A = C = D = I. Then, the matrix inversion lemma yields ( I + B)−1 = I − B( I + B)−1

(A.45)


Appendix B: Quadratic Functions, Important Derivatives, Fourier Integrals, and Parseval’s Relation B.1 QUADRATIC FUNCTIONS With a single variable x1, the quadratic function is simply I = αx12

(B.1)

With two variables x1 and x2 , the quadratic function is I = αx12 + βx22 + γx1 x2

(B.2)

Equation B.2 can be expressed as I = x T Px

(B.3)

where x = ⎡⎣ x1

x2 ⎤⎦

T

(B.4)

and P is a nonunique matrix; for example, ⎡α P=⎢ ⎣γ

0⎤ ⎥ β⎦

or

⎡α P=⎢ ⎣2 γ

−γ ⎤ ⎥ β⎦

or

⎡ α P=⎢ ⎣γ / 2

γ / 2⎤ ⎥ β ⎦

(B.5)

Out of three examples of P shown in (B.5), note that one of them is a symmetric matrix. 417


418


A quadratic function in n variables can always be written as Equation B.3 where x = ⎡⎣ x1

x2

.

.

x n ⎤⎦

T

(B.6)

and P is an n × n matrix that can always be chosen to be symmetric. Usually, the matrix P is chosen to be symmetric without any loss of generality. In fact, Equation B.3 can always be expressed as T ⎛ ⎞ T P−P ⎟x + x ⎜ ⎠ ⎝ 2

⎛ P + PT I = xT Px = xT ⎜ ⎝ 2

⎞ ⎟x ⎠

(B.7)

The matrix ( P + P T ) / 2 is symmetric, whereas the matrix ( P − P T ) / 2 is skew symmetric. For any skew-symmetric matrix, ⎛ P − PT xT ⎜ ⎝ 2

⎞ ⎟x = 0 ⎠

(B.8)

P + PT )x 2

(B.9)

Therefore, from (B.7),

I = x T Px = x T (

Definition B.1: Positive Definite Quadratic Function For a positive definite quadratic function I, I = x T Px > 0

when

x≠0

(B.10)

Definition B.2: Positive Semidefinite Quadratic Function For a positive semidefinite quadratic function I, I = x T Px ≥ 0

when

x≠0

(B.11)

As explained earlier, the matrix P is chosen to be symmetric in (B.10) and (B.11) without any loss of generality; i.e., P = P T . Fact B.1 All eigenvalues of a symmetric matrix are real, and we can always find n orthogonal eigenvectors irrespective of whether eigenvalues are distinct or repeated (Strang, 1988).


Appendix B

419

Definition B.3: Positive Definite Matrix For a positive definite matrix P, x T Px > 0

when

x≠0

(B.12)

A positive definite symmetric matrix is denoted as P = PT > 0

(B.13)

Fact B.2 All eigenvalues of a positive definite symmetric matrix P = P T are positive and nonzero numbers (Strang, 1988). Definition B.4: Positive Semidefinite Matrix For a positive semidefinite matrix P, x T Px ≥ 0

when

x≠0

(B.14)

A positive semidefinite symmetric matrix is denoted as P = PT ≥ 0

(B.15)

Fact B.3 All eigenvalues of a positive semidefinite symmetric matrix P = P T are zero or positive numbers (Strang, 1988).

B.2 DERIVATIVE OF A QUADRATIC FUNCTION Fact B.5 Let I = x T Px ;

P = PT

(B.16)

Then, ∂I ⎡ ∂I = ∂x ⎢⎣ ∂x1

∂I ∂x2

T

.

.

∂I ⎤ = 2 Px ∂x n ⎥⎦

(B.17)


420


Proof Denote ⎡ p11 ⎢ ⎢ p21 P=⎢ . ⎢ ⎢ . ⎢p ⎣ n1

p12 p22 . . pn 2

. . . . .

. . . . .

p1n ⎤ ⎥ p2 n ⎥ . ⎥ ⎥ . ⎥ pnn ⎥⎦

(B.18)

The condition P = P T is expressed as pij = p ji

(B.19)

Equation B.16 can be written as I = x1 [ p11x1 + p12 x2 + … + p1n xn ] +

x2 [ p21x1 + p22 x2 + … + p2 n xn ]

+

x3 [ p31x1 + p32 x2 + … + p3n xn ]

(B.20)

. . +

xn [ pn1x1 + pn 2 x2 + … + pnn xn ]

Therefore, ∂I = 2 p11x1 + ( p12 x2 + … + p1n xn ) + p21x2 + … pn1xn ∂x1

(B.21)

Because of (B.19), ∂I = 2( p11x1 + p12 x2 + … + p1n xn ) ∂x1

(B.22)

∂I = 2( pi1x1 + pi 2 x2 + … + pin xn ) ∂xi

(B.23)

In general,


Appendix B

421

Therefore, ⎡ ∂I ⎤ ⎢ ∂x ⎥ ⎢ 1⎥ ⎡ p11 ⎢ ∂I ⎥ ⎢ ⎢ ⎥ ⎢ p21 ∂ x ⎢ 2 ⎥=2⎢ . ⎢ . ⎥ ⎢ ⎢ ⎥ ⎢ . ⎢ . ⎥ ⎢p ⎣ n1 ⎢ ∂I ⎥ ⎢ ⎥ ⎢⎣ ∂x n ⎥⎦

p12 p22 . . pn 2

. . . . .

. . . . .

p1n ⎤ ⎡ x1 ⎤ ⎥⎢ ⎥ p2 n ⎥ ⎢ x2 ⎥ . ⎥⎢ . ⎥ ⎥⎢ ⎥ . ⎥⎢ . ⎥ pnn ⎥⎦ ⎢⎣ x n ⎥⎦

(B.24)

or ∂I = 2 Px ∂x

(B.25)


B.3 DERIVATIVE OF A LINEAR FUNCTION Fact B.6 Let J = λT x

(B.26)

where λ = [ λ1

λ2

.

.

λn ]

T

(B.27)

Then, ∂J ⎡ ∂J =⎢ ∂x ⎣ ∂x1

∂J ∂x2

∂J ⎤ ⎥ =λ ∂xn ⎦ T

.

.

(B.28)

Proof From (B.26), J = λ1x1 + λ 2 x2 + … + λ n xn

(B.29)


422


Then, ∂J = λi ; ∂xi ∂J ⎡ ∂J =⎢ ∂x ⎣ ∂x1

∂J ∂x2

i = 1, 2, …, n

∂J ⎤ ⎥ = [ λ1 ∂xn ⎦

(B.30)

T

.

.

λ2

.

.

λ n ]T = λ

(B.31)


B.4 FOURIER INTEGRALS AND PARSEVAL’S THEOREM B.4.1 SCALAR SIGNAL Consider a signal x (t )us (t ) ∈ L2 where the L2 space is defined in Chapter 5, Section 5.5.1, and us (t ) is the unit step function defined as ⎧1 us (t ) = ⎨ ⎩⎪0

t≥0 t0.

(B.39)

where the unit step function us (t ) is defined by (B.32). The Laplace transform of the signal (B.39) is

x (s ) =

b (s + a )2 + b 2

(B.40)

Because x (t ) ∈ L2, the Fourier transform can be obtained by substituting s = jω in (B.40):

x ( jω ) =

b ( j ω + a )2 + b 2

(B.41)


424


Scalar Parseval Relation ∞

∫

| x (t ) |2 dt =

0

1 2π

+∞

∫ | x ( jω ) | d ω 2

(B.42)

−∞

where x ( jω ) is the Fourier transform of x (t ) ∈ L2 . Proof Using (B.35), ∞

∫

∞

| x(t ) |2 dt =

0

∞

∫

∫

x(t )x * (t )dt =

0

0

⎡ 1 x(t )⎢ ⎢⎣ 2 π

⎤ x( jω)e jωt dω⎥ * dt ⎥⎦ −∞

+∞

∫

(B.43)

Interchanging the order of integrations on the right side of (B.43), ∞

∫ 0

1 | x(t ) | dt = 2π 2

⎤ ⎡ +∞ x ( jω)⎢ x(t )e− jωt dt ⎥ dω ⎥⎦ ⎣⎢ 0 −∞ ∞

∫

*

∫

(B.44)

Using the definition (B.33), ∞

∫

| x(t ) |2 dt =

0

1 2π

∞

∫

x *( jω)x( jω)dω =

−∞

1 2π

∞

∫ | x ( j ω) |

2

dω

(B.45)

−∞


B.4.2 VECTOR SIGNAL For an n-dimensional vector signal x(t )us (t ) ∈ L2 with the unit step function us (t ) defined by (B.32), the Fourier transform pair is defined as ∞

x ( jω ) =

∫ x(t)e

− jωt

dt

(B.46)

0

and 1 x (t ) = 2π

∞

∫ x( jω)e

−∞

jωt

dω

(B.47)


Appendix B

425

Multivariable Parseval Relation For a symmetric matrix Q, ∞

∫ 0

1 x (t )Qx (t ) dt = 2π H

+∞

∫x

H

( jω )Qx ( jω ) d ω

(B.48)

−∞

where x( jω ) is the Fourier transform of x(t ) ∈ L2 . Proof Using (B.47), ∞

∫x

∞

H

(t )Qx (t ) dt =

0

+∞

∫ ∫ [x( jω)e 0

1 2π

jωt

d ω ]H Qx (t ) dt

(B.49)

x H ( jω )Q e − jωt x (t ) dtdω

(B.50)

−∞

Interchanging the order of integrations in (B.49), ∞

∫

x H (t )Qx (t ) dt =

0

1 2π

∞

∫

∞

∫

−∞

0

Using (B.46), ∞

∫ 0

1 x (t )Qx (t ) dt = 2π H


∞

∫x

−∞

H

( jω )Qx ( jω ) d ω

(B.51)



Appendix C: Norms, Singular Values, Supremum, and Infinimum C.1 VECTOR NORMS For a complex or real vector, the norm . is a real valued function with the following properties (Vidyasagar, 1993): i.

x > 0 for all x ≠ 0

(C.1)

ii.

x = 0 implies x = 0

(C.2)

iii.

αx = α x for real or complex number α

(C.3)

iv.

x + y ≤ x + y for any complex or real vectors x and y

(C.4)

Three commonly used norms for a complex or real vector x are defined as p

p

p

x p = ( x1 + x2 + … xn )1/ p

where p = 1, 2, and ∞

(C.5)

Note that x

∞

= max xi i

(C.6)

C.2 MATRIX NORMS For a vector norm, the induced norm of a real or complex matrix A is defined as

A

p

= sup x ≠0

Ax x

p p

= sup Ax

(C.7)

x =1

427


428


It has been shown (Vidyasagar, 1993) that A = max 1

j

∑a

;

ij

maximum column sum

(C.8)

maximum row sum

(C.9)

i

A

∞

= max i

∑a

ij

;

j

A = max(λ i ( A H A)1/ 2 ) 2

i

(C.10)

where A H is the complex conjugate transpose of the matrix A, and λ i ( A H A) is the ith eigenvalue of A H A . Fact C.1 The matrix A H A is positive semidefinite. Proof For a vector x, the quadratic form is q = x H A H Ax

(C.11)

y = Ax

(C.12)

q = yH y ≥ 0

(C.13)

Let

Then,

Because of this fact, eigenvalues of A H A are guaranteed to be nonnegative. EXAMPLE ⎡3 ⎢ A = ⎢ −1 ⎢1 ⎣ Here,

−3 3 2

−1⎤ ⎥ −1⎥ 2 ⎥⎦

(C.14)


Appendix C

429

A =8 , A 1

∞

= 7 and A = 5.33 2

C.3 SINGULAR VALUES OF A MATRIX Singular values σ i ( A) of a real or complex mxn matrix A are nonnegative square roots of eigenvalues of A H A . The dimension of A H A is n × n. Therefore, there will be n eigenvalues of A H A . However, if the rank of the matrix A H A is r , n–r eigenvalues of A H A will be zero. Also, note that r≤p

where

p = min(m, n )

(C.15)

Therefore, the number of singular values is chosen to be p, which is the maximum rank that a matrix A can have. Then, if there is any singular value which is zero, the matrix will not be of full rank. These singular values are usually arranged in descending order. In summary, singular values of a real or a complex matrix A are given as σ i ( A) = λ i ( A H A) ;

i = 1, 2,…, p

σ 1 ( A) ≥ σ 2 ( A) ≥ … ≥ σ p ( A)

(C.16) (C.17)

Maximum and minimum singular values are also denoted by σ and σ , respectively. Therefore,

σ( A) = max x ≠0

Ax x

2

= A

(C.18)

2

2

and ⎡ A −1x σ( A) = ⎢ max ⎢ x ≠0 x 2 ⎣

−1

2

⎤ −1 ⎥ = ⎡ A −1 ⎤ ⎢⎣ ⎥ 2⎦ ⎥ ⎦

if A −1 exists

(C.19)

The minimum singular value, σ( A) , is a measure of how near the matrix A is to being singular or rank deficient. The maximum singular value, σ( A) , is also the 2norm of the matrix A . 2

C.4 SINGULAR VALUE DECOMPOSITION (SVD) For an m × n matrix A, the SVD of the matrix A is given (Zhou et al., 1996) by


430


A = U ΣV H

(C.20)

where U and V are m × m and n × n unitary matrices, respectively; i.e., UU H = I m

VV H = I n

and

(C.21a,b)

These results imply that the 2-norms of columns of U and V are unity. Furthermore, column vectors of U are orthogonal with respect to each other. Similarly, column vectors of V are orthogonal with respect to each other. Denote U = ⎡⎣u1

u2

.

.

u m ⎤⎦

(C.22)

V = ⎡⎣v1

v2

.

.

v n ⎤⎦

(C.23)

where u i and v i are ith column vectors of matrices U and V, respectively. The matrix Σ is an mxn matrix with the following structures: Σ = ⎡⎣ Σ1

0 ⎤⎦

if m < n

(C.24)

⎡Σ ⎤ Σ = ⎢ 1⎥ ⎣0⎦

if m > n

(C.25)

Σ = Σ1

if m = n

(C.26)

where Σ1 = diag (σ1

COMPUTATION

OF

U

AND

σ2

.

.

σp)

(C.27)

V

Postmultiplying (C.20) by V, AV = U Σ

(C.28)

U H A = ΣV H

(C.29)

Premultiplying (C.20) by U H ,


Appendix C

431

Premultiplying (C.28) by A H , A H AV = A HU Σ = V Σ H Σ

(C.30)

A H Av i = σ 2i v i

(C.31)

or

Therefore, columns of V are unit (right) eigenvectors of A H A . They are called right singular vectors of the matrix A. Postmultiplying (C.29) by A H , U H AA H = ΣV H A H = ΣΣ HU H

(C.32)

u iH AA H = σ i2 u iH

(C.33)

or

Therefore, rows of U H are unit (left) eigenvectors of AA H . They are called left singular vectors of the matrix A.

SIGNIFICANCE OF SINGULAR VECTORS CORRESPONDING MAXIMUM AND MINIMUM SINGULAR VALUES

TO THE

Let v and u be the right and left singular vectors corresponding to the maximum singular value, σ , respectively. Similarly, let v and u be the right and left singular vectors corresponding to the minimum singular value, σ , respectively. As singular values are arranged in descending order, σ = σp

and

σ = σ1

(C.34)

u = up

and

u = u1

(C.35)

v = vp

and

v = v1

(C.36)

Av = σ.u

and

Av = σ u

With these notations, (C.37a,b)

If we think of a system with v i and σ i u i as the input and output, respectively, the 2-norms of the output and input are σ i and 1, respectively. Therefore, the


432


vector v ( v ) represents the input direction leading to the maximum (minimum) magnitude of the output, which occurs along the u(u) direction (Maciejowski, 1989; Zhou et al., 1996).

C.5 PROPERTIES OF SINGULAR VALUES (CHIANG AND SAFANOV, 1992)

1. σ( A) = max x ≠0

2. σ( A) = min x ≠0

Ax x

= A

2

(C.38)

2

Ax x

2

2

(C.39)

2

3. If A −1 exists, σ ( A) =

1 1 and σ ( A) = −1 σ( A ) σ ( A −1 )

(C.40)

4. σ (αA) = α σ ( A) for a scalar α

(C.41)

5. σ ( A + B) ≤ σ ( A) + σ ( B)

(C.42)

6. σ ( AB) ≤ σ ( A).σ ( B)

(C.43)

7. max[0, (σ ( A) − σ ( B))] ≤ σ ( A + B) ≤ σ ( A) + σ ( B)

(C.44)

8. σ ( A) ≤ λ( A) ≤ σ ( A) where λ( A) is an eigenvalue of A

(C.45)

n

9.

∑σ

2 i

= trace( A H A)

(C.46)

i =1

C.6 SUPREMUM AND INFINIMUM (REDDY, 1986) Let a and b be two real numbers such that b > a . Then, we define closed and open intervals as follows: i. closed interval [ a, b ] = {x : x ∈ R, a ≤ x ≤ b} It is clear that maximum value of x = b minimum value of x = a

(C.47)


Appendix C

ii. open interval ( a, b ) = {x : x ∈ R, a < x < b}

433

(C.48)

Here, it is not possible to precisely define the maximum and minimum values of x. But, upper and lower bounds can be defined. Any number greater than b is an upper bound of x. However, b is the least upper bound and is called the supremum (sup) of x. Similarly, any number smaller than a is a lower bound of x. However, a is the greatest lower bound and is called the infinimum (inf) of x.



Appendix D: Stochastic Processes A stochastic process x (t ) refers to an ensemble of functions (Papoulis, 1965), x1 (t ), x2 (t ) , x3 (t ) , …, Figure D.1. For the purpose of contents in this book, the independent variable for these functions is restricted to be time t. Therefore, the sample number in Figure D.1 can be interpreted as time instants. At any time t = ti , there will be different values of these functions, x1 (ti ) , x2 (ti ) , x3 (ti ), …, which can be plotted as a histogram. As the number of these function approaches infinity, this histogram will turn into a probability density function. In other words, there is a random variable x (ti ) and the corresponding probability density function for each time ti .

D.1 STATIONARY STOCHASTIC PROCESS A stochastic process x (t ) is said to be stationary if the probability distribution function of x (ti ) is independent of ti ; i.e., it is time invariant. The mean value and standard deviation of x (ti ) are also time invariant. The mean or expected value of x (t ) at any time t is given by ∞

E ( x (t )) =

∫ x(t) p( x(t))dx = a constant value, η

x

(D.1)

−∞

where p ( x ) is the probability density function. For a stationary and Gaussian stochastic process, the probability density function is given by

p( x ) =

1 σ 2π

−

e

( x − η x )2 2 σ x2

(D.2)

where σ x is the standard deviation; i.e., σ x 2 = E (( x − η)2 ) = E ( x 2 ) − ηx 2

(D.3)

435


436


1 +

x

x

0.9

x

+

+

x x

0.8

+ +

x

+

x x

0.7

x

x

+

+

x

+

+

+

+

x

+

0.6

+

x x

+

x

+

0.5

+

x

+

+

0.4

x

x

x x

+x

x

x

x

0.2

+

x

0.3

+ +

+ + Member 1

+

x

+

+

x Member 2

0.1

Member 3

+ 0 0

+ x

x

+ 5

x

x

10

15 Sample Number

20

25

30

FIGURE D.1 Ensemble members of a stochastic process.

and ∞

E (x2 ) =

∫ x p( x)dx 2

(D.4)

−∞

AUTOCORRELATION FUNCTION The autocorrelation function of a stationary stochastic process x (t ) is defined as R xx (τ) = E ( x (t ) x (t + τ))

(D.5)

It should be noted that the autocorrelation function is only a function of τ and is independent of t. It should also be noted that R xx (0 ) = E ( x 2 (t ))

(D.6)

R xx (τ) = R xx (− τ)

(D.7)


Appendix D

437

Numerically, R(τ) will be computed from N observations of the stochastic process by 1 R xx (τ) = N

N

∑ x (t) x (t + τ) i

i

(D.8)

i =1

AUTOCOVARIANCE FUNCTION The autocovariance function of a stationary stochastic process x (t ) is defined as C xx (τ) = E (( x (t ) − ηx )( x (t + τ) − ηx ))

(D.9)

C xx ( τ ) = Rxx ( τ ) − η2x

(D.10)

C xx (0 ) = σ 2x

(D.11)

Therefore,

and

CROSSCORRELATION FUNCTION The crosscorrelation function for two different stationary stochastic processes x (t ) and y (t ) is defined as R xy (τ) = E ( x (t ) y (t + τ))

(D.12)

It should be noted that the crosscorrelation function is only a function of τ and is independent of t. Numerically, R xy (τ) will be computed from N observations of the stochastic process by 1 R xy (τ) = N

N

∑ x (t)y (t + τ) i

i

(D.13)

i =1

CROSSCOVARIANCE FUNCTION The crosscovariance function for two different stationary stochastic processes x (t ) and y (t ) is defined as C xy (τ) = E (( x (t ) − ηx )( y (t + τ) − ηy ))

(D.14)


438


Uncorrelated Stochastic Processes Two stationary stochastic processes are uncorrelated if R xy (τ) = E ( x (t ) y (t + τ)) = E ( x (t )) E ( y (t + τ)) = ηx ηy

(D.15)

Equivalently, two stationary stochastic processes are uncorrelated if C xy ( τ ) = E (( x(t ) − ηx )( y(t + τ ) − ηy )) = E (( x(t ) − ηx ))E (( y(t + τ ) − ηy )) = 0

(D.16)

A stationary stochastic process is called ergodic if all of its statistics can be determined from a single observation or a single member of the ensemble. Therefore, the mean, variance, autocorrelation function, and crosscorrelation function of ergodic processes can be computed as follows: 1 T →∞ 2T

ηx = E ( x(t )) = lim

1 T →∞ 2T

σ 2x = E (( x(t ) − η)2 ) = lim

+T

∫ x(t )dt

−T

+T

∫ (x(t ) − η) dt 2

(D.18)

−T

1 T →∞ 2T

Rxx ( τ ) = E ( x(t )x(t + τ )) = lim

1 T →∞ 2T

Rxy ( τ ) = E ( x(t )y(t + τ )) = lim

(D.17)

+T

∫ x(t )x(t + τ)dt

(D.19)

−T

+T

∫ x(t )y(t + τ)dt

(D.20)

−T

D.2 POWER SPECTRUM OR POWER SPECTRAL DENSITY (PSD) The PSD, S xx (ω ) , of a stationary stochastic process x (t ) is the Fourier transform of its autocorrelation function: ∞

S xx (ω ) =

∫e

−∞

− jωτ

R xx (τ) d τ

(D.21)


Appendix D

439

The inverse Fourier transform corresponding to Equation D.21 yields 1 R xx (τ) = 2π

∞

∫e

jωτ

S xx (ω ) d ω

(D.22)

−∞

D.3 WHITE NOISE: A SPECIAL STATIONARY STOCHASTIC PROCESS The autocorrelation function of the white noise, v (t ) , is given by Rvv (τ) = Pδ(τ)

(D.23)

where P is a constant, and δ(τ) is a delta-Dirac function. Equation D.23 indicates that Rvv (τ) = 0

when

τ≠0

(D.24)

Therefore, there is no correlation between v (t1 ) and v (t2 ) where t1 ≠ t2 . Because Rvv (0 ) is infinite, the standard deviation of the process is infinite. The PSD of the white noise, v (t ), is given by ∞

S vv (ω ) =

∫e

− jωτ

P δ (τ) d τ = P

(D.25)

−∞

That is, the PSD of white noise is a constant for each frequency, and consequently, white noise will have infinite power. Thus, white noise is a mathematical fiction.

GENERATION OF APPROXIMATE GAUSSIAN WHITE NOISE IN MATLAB Gaussian random numbers with the zero mean and the standard deviation σ can be generated as follows: w ( k ) = σr ( k );

k = 1, 2, 3, …

(D.26)

where random numbers r(k) wth zero mean and unity standard deviation are generated using the Matlab routine randn. Then, a time-domain signal is obtained by considering this sequence of numbers w(k) as a discrete-time signal with a constant sampling interval T. The continuous-time signal v(t) obtained via a zero-order hold as shown in Figure D.2 can be viewed as a Gaussian white noise with the power spectral density σ 2 T (Meditch, 1969).


440


v (t ) w(5) w(3) w( 4 )

w( 2 )

w(6)

w(1)

0

T

3T

2T

t

5T

4T

FIGURE D.2 Continuous-time white noise v(t) from random numbers w(k). x (t )

y (t ) G (s )

FIGURE D.3 A SISO system.

D.4 RESPONSE OF A SISO LINEAR AND TIME-INVARIANT SYSTEM SUBJECTED TO A STATIONARY STOCHASTIC PROCESS Consider the linear system, Figure D.3, with the transfer function G (s ) where the input x(t) is a stationary stochastic process with PSD equal to S xx (ω ) . The output y (t ) will also be a stochastic process. For a stable system, the output will become a stationary stochastic process in the statistical steady state. In the statistical steady state, 2

S yy (ω ) = G ( jω ) S xx (ω )

(D.27)

where S yy (ω ) is the PSD of the output y (t ) . When the input is a white noise with unit PSD, i.e., S xx (ω ) = 1 , S yy (ω ) = G ( jω )

2

(D.28)

D.5 VECTOR STATIONARY STOCHASTIC PROCESSES Define the vector v(t ) = ⎡⎣ v1 (t )

v2 (t )

.

.

v n (t ) ⎤⎦

T

(D.29)

where v1 (t ) , v2 (t ) , …, v n (t ) are n scalar stochastic processes. Then, the following quantities are defined (Athans, 1992; Kwakernaak and Sivan, 1972):


Appendix D

441

Mean η = E (v (t )) = [ E (v1 (t ))

[η1

E (v2 (t )) η2

.

.

.

.

E (vn (t ))] =

ηn ]

T

(D.30)

T

Autocorrelation Matrix R (τ) = E (v (t )(v (t + τ))T )

(D.31)

C ( τ ) = E ((v (t ) − η)(v (t + τ ) − η)T )

(D.32)

Ξ = C (0 ) = E ((v (t ) − η)(v (t ) − η)T )

(D.33)

Autocovariance Matrix

Variance Matrix

PSD Matrix ∞

S (ω ) =

∫e

− jωτ

R (τ) d τ

(D.34)

−∞

R (τ) =

1 2π

∞

∫e

−∞

jωτ

S (ω ) d ω

(D.35)



Appendix E: Optimization of a Scalar Function with Constraints in the Form of a Symmetric Real Matrix Equal to Zero (Zhou et al., 1996) Definition E.1 Suppose that an n-dimensional real vector x o is a local minimum of the scalar function f ( x ) subject to constraints U ( x ) = 0 where U ( x ) is a symmetric real matrix. Then, x o is a regular point of the constraints if for a symmetric matrix P, ∇trace(U ( x o ) P ) = 0 has a unique solution P = 0 . Definition E.2 Let x = ⎡⎣ x1

x2

.

T

x n ⎤⎦ . Then, the gradient ∇f ( x ) is defined as

.

⎡ ∂f ∇f ( x ) = ⎢ ⎣ ∂x1

∂f ∂x2

.

.

∂f ⎤ ∂x n ⎥⎦

T

(E.1)

Definition E.3 Given an m × n matrix X, define the following vector: x = VecX = ⎡⎣ x11

.

xm 1

x12

.

xm 2

.

x1n

.

xmn ⎤⎦

T

Then, for a scalar function f ( x ) , the gradients with respect to the matrix X and x = VecX are defined as

443


444


⎡ ∂f ( x ) ⎢ ∂x ⎢ 11 ⎢ . ∂f ( x ) ⎢ =⎢ . ∂X ⎢ . ⎢ ⎢ ∂f ( x ) ⎢ ⎣ ∂xm 1

∂f ( x ) ∂x12

.

.

.

.

.

. . ∂f ( x ) ∂xm 2

. .

. .

.

.

∂f (x ) xm 2

.

∂f ( x ) ⎤ ∂x1n ⎥⎥ . ⎥ ⎥ . ⎥ . ⎥ ⎥ ∂f ( x ) ⎥ ∂xmn ⎥⎦

(E.2)

∇f (x ) = ⎡ ∂f (x ) ⎢ ⎣ x11

.

∂f (x ) x m1

∂f (x ) x12

.

∂f (x ) x1n

.

∂f (x ) ⎤ (E.3) ⎥ xmn ⎦

It should be noted that ∂f ( x ) =0 ∂X

(E.4)

∇f ( x ) = 0

(E.5)

is equivalent to

Fact E.1 Suppose that an n-dimensional real vector x o is a local minimum of the scalar function f ( x ) subject to constraints U ( x ) = 0 where U ( x ) is a symmetric real matrix. Also, assume that x o is a regular point of the constraints. Then, there exists a unique multiplier symmetric matrix P such that if we define g ( x, P ) = f ( x ) + trace(U ( x ) P )

(E.6)

then at the optimal solution x o , Po , ∇g ( x 0 , Po ) = ∇f ( x 0 ) + ∇trace(U ( x 0 ) Po ) = 0

(E.7)

Note that the gradient here is defined with respect to the following vector: ⎡ x ⎤ ⎢ ⎥ ⎣ vecP ⎦

(E.9)


Appendix E

445

Fact E.2 ∂ tr ( AXB) = AT BT ∂X

(E.10)

∂ tr ( AX T B) = BA ∂X

(E.11)

∂ tr ( AXBX ) = AT X T BT + BT X T AT ∂X

(E.12)

∂ tr ( AXBX T ) = AT XBT + AXB ∂X

(E.13)



Appendix F: A Flexible Tetrahedral Truss Structure The tetrahedral flexible structure, Figure F.1, has been developed at the C. S. Draper Laboratory (Strunce et al., 1979). The structure is supported on the ground by three right-angled bipods on which actuators are mounted. The finite element model is represented by + Kp(t ) = B f f (t ) Mp

(F.1)

where M and K are mass and stiffness matrices, respectively. The vector p(t) contains displacements in x, y, and z directions of nodes 1-4. The 3 × 1 force vector f (t ) is composed of the forces provided by actuators mounted on elements 1 -3. The modal matrix Φ and the diagonal eigenvalue matrix Λ are defined as follows: K Φ = M ΦΛ

(F.2)

where Λ = diag[ω12

ω 22

.

.

2 ω12 ]

(F.3)

and ΦT M Φ = I

(F.4)

The modal matrix is provided in the report (Strunce et al., 1979) and the thesis (Kao, 1989). The natural frequencies are listed in Table F.1.

447


448


Z

1

12

11

10

4

8

2

1

3

9 2 10

3

y

9

5

6

7 x

8

4

7

5 6

FIGURE F.1 A flexible tetrahedral truss structure.

TABLE F.1 Natural Frequencies (rad/sec) ω1

ω2

ω3

ω4

ω5

ω6

ω7

ω8

ω9

ω10

ω11

ω12

1.342

1.643

3.391

4.207

4.658

4.754

8.538

4.754

8.538

9.252

10.296

12.923

Define modal coordinate vector ψ as p(t ) = Φψ ψ(t )

(F.5)

Substituting (F.5) into (F.1) and premultiplying both sides by ΦT , + Λψ(t ) = Bψ f (t ) ψ where

(F.6)


Appendix F

449

⎡ −0.023 ⎢ ⎢ −0.112 ⎢ −0.077 ⎢ ⎢ 0.189 ⎢ 0.156 ⎢ ⎢ −0.289 T Bψ = Φ B f = ⎢ ⎢ −0.320 ⎢ 0.365 ⎢ ⎢ −0.229 ⎢ 0.167 ⎢ ⎢ −0.145 ⎢ 0.025 ⎣ ψT = [ ψ1

ψ2

−0.067 0.017 0.271 −0.060 −0.049 0.289 −0.369 0.299 0.250 −0.150 0.146 −0.013 .

.

−0.044 ⎤ ⎥ 0.069 ⎥ 0.046 ⎥ ⎥ −0.249 ⎥ 0.351 ⎥ ⎥ −0.289 ⎥ ⎥ −0.049 ⎥ −0.069 ⎥ ⎥ 0.231 ⎥ −0.317 ⎥ ⎥ −0.220 ⎥ 0.114 ⎥⎦ ψ12 ]

(F.7)

(F.8)



Appendix G: Space Shuttle Dynamics during Reentry (Doyle et al., 1986) The 4-state rigid body aircraft model (Figure G.1) has state variables x(t) and measurements y meas : ⎡ β ⎤ ⎡sideslip ⎢ ⎥ ⎢ p roll x=⎢ ⎥=⎢ ⎢ r ⎥ ⎢ ya aw ⎢ ⎥ ⎢ ⎢⎣ φ ⎥⎦ ⎢⎣ bank

angle ⎤ ⎥ rate ⎥ ; rate ⎥ ⎥ angle ⎥⎦

y meas

⎡p⎤ ⎢ ⎥ r =⎢ ⎥ ⎢ ηy ⎥ ⎢ ⎥ ⎢⎣ φ ⎥⎦

(G.1a,b)

where ηy is the lateral acceleration. In Figure G.1, α and V are the angle of attack and the velocity of the shuttle, respectively. Also,

c fm

⎡ c y ⎤ ⎡ side ⎢ ⎥ ⎢ = ⎢c n ⎥ = ⎢ yawing ⎢ c ⎥ ⎢ rolling ⎣ l⎦ ⎣

⎡ c β ⎤ ⎡sideslip force ⎤ ⎢ ⎥ ⎢ ⎥ moment ⎥ and c an = ⎢c α ⎥ = ⎢ aileron ⎢ c ⎥ ⎢ rudder moment ⎥⎦ ⎢⎣ l ⎦⎥ ⎣

angle ⎤ ⎥ angle ⎥ (G.2a,b) angle ⎥⎦

Then, c fm = C aer c an

C aer

⎡ c yβ ⎢ = ⎢c nβ ⎢ ⎣ clβ

c ya c na cla

(G.3) c yr ⎤ ⎥ c nr ⎥ ⎥ clr ⎦

(G.4)

451


452


Y rudder

elevon

X roll, p Center of Mass

Vx Vz

yaw, r

bank angle Vy Z

V

FIGURE G.1 Space shuttle variables (Doyle et al., 1986).

rp c fm

c an

R

u act

Wact e act

ny rp

r 0.037 e perf W perf

Space Shuttle

u act

ideal

y meas

gust noise

Ideal Model

Wsens

controller com

FIGURE G.2 Space shuttle control system (Doyle et al., 1986).


Appendix G

453

The elements of the aerodynamic coefficients matrix C aer are uncertain. Let C aer = C aer + δC aer

(G.5)

where

δC aer

⎡ryβ ⎢ R=⎢0 ⎢ ⎣0 ⎡δ yβ ⎢ Δ =⎢ 0 ⎢ ⎣ 0 T

⎡ ryβ δ yβ ⎢ = ⎢rnβ δ yβ ⎢ ⎣ rlβ δ lβ

rya δ ya rna δ na rla δ la

ryr δ yr ⎤ ⎥ rnr δ nr ⎥ = R Δ ⎥ rlr δ lr ⎦

(G.6)

0⎤ ⎥ 0⎥ ⎥ rlr ⎦

0

0

rya

0

0

ryr

0

rnβ

0

0

rna

0

0

rnr

0

rlβ

0

0

rla

0

0

0

0

δ ya

0

0

δ yr

0

δ nβ

0

0

δ na

0

0

δ nr

0

δ lβ

0

0

δ la

0

0

0⎤ ⎥ 0⎥ ⎥ δ lr ⎦

(G.7)

(G.8)

Let c fm = c fm + δc fm

(G.9)

δc fm = δC aer c an = R Δc an

(G.10)

x = Ax (t ) + Bu(t )

(G.11)

y (t ) = Cx (t ) + Du(t )

(G.12)

where c fm is the nominal value. Then,

STATE SPACE MODEL

where


454


⎡δc fm ⎤ ⎢ ⎥ u = ⎢ u act ⎥ ; ⎢ gust ⎥ ⎣ ⎦

⎡ c ⎤ y = ⎢ an ⎥ ; ⎣ y meas ⎦

⎡ −9.460 e − 02 ⎢ −3.5595e + 00 A=⎢ ⎢ 3.950 e − 01 ⎢ 0 ⎢⎣

1.409e − 01 −4.284e − 01 −1.263e − 02 1

and

⎡u ⎤ u act = ⎢ e ⎥ ⎣ur ⎦

−9.900 e − 01 2.809e − 01 −8.142 e − 02 −1.405e − 01

(G.13)

3.637e − 02 ⎤ ⎥ 0 ⎥ (G.14) ⎥ 0 ⎥ 0 ⎥⎦

B= ⎡1.275 e − 05 ⎢ 0 ⎢ ⎢ 0 ⎢ 0 ⎣

0 −3.114 e − 05 −1.905 e − 04

0 −3.117 e − 03 −6.443e − 05

−1.240 e − 02 6.571e + 00 3.783e − 01

1.023e − 02 1.256 e + 00 −2.560 e − 01

0

0

0

0

−1.086 e − 04 ⎤ ⎥ −4.126 e − 03⎥ 4.533e − 04 ⎥ ⎥ 0 ⎦

(G.15) ⎡ 1 ⎢ 0 ⎢ ⎢ 0 ⎢ C=⎢ 0 ⎢ 0 ⎢ ⎢ −6.804e + 01 ⎢ 0 ⎣

0 0 0 1 0 −1.744 0

0 0 0 0 1 −4.058 0

⎤ 0 ⎥ 0 ⎥ ⎥ 0 ⎥ 0 ⎥ ⎥ 0 ⎥ −3.720 e − 05 ⎥ ⎥ 1 ⎦

(G.16)

D= ⎡ 0 ⎢ 0 ⎢ ⎢ 0 ⎢ 0 ⎢ ⎢ 0 ⎢ ⎢1.111e − 02 ⎢ 0 ⎣

0 0 0 0 0 −1.111e − 02 0

0 0 0 0 0 −1.111e − 02 0

0 1 0 0 0 2.667 e + 01 0

0 0 1 0 0 −2.952 0

1.148 e − 03 ⎤ ⎥ 0 ⎥ ⎥ 0 ⎥ 0 ⎥ ⎥ 0 ⎥ −7.810 e − 02⎥ ⎥ 0 ⎦

(G.17) Each measurement is corrupted by the white noise passing through a filter. The weighting matrix Wsens is described as follows:


Appendix G

455

Wsens = ⎡ 0.003(1 + 100 s ) diag ⎢ (1 + 2 s ) ⎣

0.003(1 + 100 s ) (11 + 2 s )

0.25(1 + 20 s ) (1 + 0.1s )

0.007(1 + 100 s ) ⎤ ⎥ (1 + 0.55 s) ⎦ (G.18)

As p and r are both measured with similar gyroscopes, their sensor noise weights are assumed to be identical. The actuators are represented by the elevon u e and the rudder u r . The actuator weights are Wact = diag ⎡⎣We

Wr ⎤⎦

(G.19)

where

We =

1 1 − 1.732 (s / 173) + (s / 173)2 1 + 1.44(s / 14) + (s / 14)2 1 + 1.732 (s / 173) + (s / 173)2

(G.20a)

Wr =

1 1 − 1.732 (s / 173) + (s / 173)2 2 1 + 1.5(s / 21) + (s / 21) 1 + 1.732 (s / 173) + (s / 173)2

(G.20b)

The performance weights are ⎡ 0.8(1 + s ) W perf = diag ⎢ ⎣ (1 + 10 s )

500 (1 + s ) (1 + 100 s )

250 (1 + s ) ⎤ (1 + 100 s ) ⎥⎦

(G.21)

The dynamics of the ideal model is given by Gideal (s ) =

1 1 + 1.4(s / 1.2 ) + (s / 1.2 )2

(G.22)



Index A Abate studies, 76 Active suspension and active suspension systems full state feedback control, 319–320 H2 control, 285–286 optimal linear state feedback, 134–137 stochastic optimal regulator, 211–212 Actuators and actuator weights, 455 Additive uncertainty, 301, 301–304 Algebraic Riccati equation (ARE), see also Control algebraic Riccati equation (CARE); Filter algebraic Riccati equation (FARE); Riccati equation double integrator system, 385 infinite final time, 132 robustness, 253 Analog computer simulation diagram, 7–8 Anderson and Moore studies, 127, 132, 203, 239–242, 257 Annaswamy studies, 231 Aortic pressure observer, 182–184 ARE, see Algebraic Riccati equation (ARE) Armature-controlled DC motor, 128–132, 129 Artificial heart, 182–184 Asada and Slotine studies, 377, 387 Astrom studies, 86 Athans, Lublin and, studies, 390 Athans studies, 217, 259–260, 268, 440 Attractive condition, 377 Autocorrelation function, 436–437 Autocorrelation matrix, 441 Autocovariance function, 437 Autocovariance matrix, 441 Autonomous system, 115–123 Available bandwidth, 345

B Balancing a pointer, 82–85, 190–191 Balas studies, 225 Banavar and Dominic studies, 233 Bang-bang control, 160, 160, 168 Bass-Gura formula, 82, 85

Benchmark robust control controller, structured singular values analysis, 360–364 modeling errors impact, 223–224, 226–228 Bernstein, Wie and, studies, 223–224 Bicycle dynamics, 86–89 Biegler studies, 117, 120 Bode’s sensitivity integrals, 345–351 Bounded real lemma, 273, 275–278, see also Lemmas Braatz and Morari studies, 360 Bryson studies, 142–143 Bucy, Kalman and, studies, 182 Butterworth configuration, 142

C Cannon and Schmitz studies, 232 Canonical form, similarity transformations, 30–31 CARE, see Control algebraic Riccati equation (CARE) Cascaded linear systems, 4 Cases controller design, 335–336 linear quadratic regulator, 132–137 minimum-time control, 158–159 small gain theorem, 295–297 Cayley-Hamilton theorem, 23, 65 Chen, Seo and, studies, 423 Chen studies, 423 Chiang and Safanov studies, 336, 338, 366, 432 Closed-loop observer aortic pressure observer, 182–184 artificial heart, 182–184 detectability, 181 disturbance observer, 184–186 fundamentals, 180, 181 multiple-output system, 182 observer gain vector determination, 181–182 observer pole locations, 182–186 proof, 181 single-output system, 181–182 theorem, 181 Closed-loops optimal control, linear quadratic regulator, 126–127

457


458


system dynamics, linear quadratic Gaussian control, 218–219 transfer function, state feedback effect, 80–89 Column operations, 50–51 Column-reduced matrix, 38 Column space, 409 Combined observer-controller, 186–191 Common right divisor, see Greatest common right divisor (GCRD) Complementary sensitivity, robustness definitions, 240 facts, 240 fundamentals, 238–242 Kalman filter, 262 Matlab programming, 247–251 multiplicative uncertainty, 246–247 proof, 240 spring-mass system, 247–251 structured uncertainties, 242, 242–244 unstructured uncertainties, 244, 244–251 Computations, H2 and H∞ norms, 278–283 Conditions, optimal control theory, 113–115 Conservation laws, 26, 345 Constant input disturbances, 91–95 Constant output, 89–91 Constraints Hamiltonian, autonomous system, 117–120 loop shaping, 345–351 optimality conditions, 114–115 Continuous-time Gaussian white noise vector, 201, 216 Control bicycle dynamics, 88–89 cost, symmetric root locus, 140–142 frequency-shaped linear quadratic control, 148–154 pointer balancing, 84 Control, estimated states active suspension system, 211–212 aortic pressure observer, 182–184 artificial heart, 182–184 benchmark robust control problem, 223–224, 226–228 closed-loop observer, 180–186, 181 closed-loop system dynamics, 218–219 combined observer-controller, 186–191 detectability, 181 disturbance observer, 184–186 double integrator system, 204–205, 212–213, 219–220 exercise problems, 228–233 flexible mode, 226–228 Kalman filter, 200–208 linear continuous-time system, 195–200 linear quadratic Gaussian control, 215–222

Matlab programming, 198–200, 205–208, 214–215, 220–221 mean, 196–197 measurement equation, 200–201, 216 modeling errors impact, 222–228 multiple-output system, 182 objective function, 209–211, 216–219 observer-based control, 222–228 observer dynamics, 201–202 observer gain matrix determination, 182 observer gain vector determination, 181–182 open-loop observer, 179–180 optimal regulator, 208–215 pointer balancing, 190–191, 191 pole locations, 182–186 proof, 181 reduced-order observer, 191–195, 193 regulator, optimal, 208–215 rigid mode, 226 robust control problem, 223–224, 226–228 simple mass, 194, 194–195, 204–205, 212–213, 219–220 single-output system, 181–182 spring-mass damper, 198 state dynamics, 200, 208, 216 steady state, 208–215 stochastic forces, 198, 204–205, 208–215 structured parametric uncertainties, 222–224 theorem, 181 unmodeled dynamics, 224–228 variance, 197 white noise response, 195–200 Control algebraic Riccati equation (CARE), 218, 264, see also Algebraic Riccati equation (ARE); Filter algebraic Riccati equation (FARE); Riccati equation Controllability matrix, 22 MIMO system, 44–46 similarity transformations, 30–31 simultaneous, 30–36 SISO system, 19–28, 25, 27 Controllability Gramian, 19 Controller, structured singular values analysis benchmark robust control, 360–364 definitions, 351–353 D-K iteration, 366–367 fundamentals, 351–353 graphical interpretation, 360 Matlab programming, 364–366 performance, 354–366 properties, 351–353 robustness analysis, 353–354 single-input, single-output system, 357–360


Index synthesis, 366–367 theorems, 353–354 Controller design, 332–336 Controller form realization, 59–66 Control spillover, 225 Conversion, controller form realization, 64–66, 65 Cost of control, 140–142 Covariance matrix, 197 Crosscorrelation function, 437 Cross term, 127–132 Crossvariance function, 437

D Dahleh studies, 355 DC motor, 128–132 Definitions controller, structured singular values analysis, 351–353 H2 and H∞ norms, 269–270, 273 matrix-fraction description, 49 MIMO system, 37–38, 44–45 optimization, scalar function with constraints, 443–444 sensitivity, robustness, 240 well-posedness and internal stability, 289 Degree of a square polynomial matrix, 47–49 Delta-Dirac function, 439 Derivatives linear function, 421–422 quadratic function, 419–421 Detectability, closed-loop observer, 181 Diagonal form, similarity transformations, 29–30 Disturbance feedforward, 320–322 Disturbance observer, 184–186 D-K iteration, 366–367 Dominic, Banavar and, studies, 233 Dorato studies, 368 Double integrator and double integrator system full state feedback, linear system, 384–385 H2 control, 285, 287–288 Kalman filter, 204–205, 262 linear quadratic Gaussian control, 219–220 loop transfer recovery, 268–269 stochastic optimal regulator, 212–213 time-optimal control, 156–160 Doyle and Stein studies, 244, 268, 382 Doyle studies, 263, 351, 451–455 Drazenovic studies, 386 Duality, state space model properties, 10 Dynamics, space shuttle reentry, 451–455

459

E Eigenvalues and eigenvectors aortic pressure observer, 184 benchmark robust control, 224 combined observer-controller, 188 diagonal form, 29–30 disturbance observer, 186 estimated states, 396 fundamentals, 415 H∞ norm, 282–283 infinite final time, 132 LQR problem, root locus plot, 139 matrix norms, 428 observability, 33–34 open-loop stability, 84 single input/single output system, 358 sliding hyperplane matrix, 382 sliding mode, 378 state feedback gain matrix, multiinput system, 99–112, 102 symmetric matrix, 418 Electric ventricular assist device (EVAD), 182–184, 183 Electronics navigation, 101–108, 102, 108–112 Elementary column/row operations, 50–51 Energy control, minimum, 128–132 Equations solutions, state space description, 12–14 Equivalence, H∞ control, 325, 328–332 Equivalent control, sliding mode, 378 Estimated states, control active suspension system, 211–212 aortic pressure observer, 182–184 artificial heart, 182–184 benchmark robust control problem, 223–224, 226–228 closed-loop observer, 180–186, 181 closed-loop system dynamics, 218–219 combined observer-controller, 186–191 detectability, 181 disturbance observer, 184–186 double integrator system, 204–205, 212–213, 219–220 exercise problems, 228–233 flexible mode, 226–228 Kalman filter, 200–208 linear continuous-time system, 195–200 linear quadratic Gaussian control, 215–222 Matlab programming, 198–200, 205–208, 214–215, 220–221 mean, 196–197 measurement equation, 200–201, 216 modeling errors impact, 222–228 multiple-output system, 182


460


objective function, 209–211, 216–219 observer-based control, 222–228 observer dynamics, 201–202 observer gain matrix determination, 182 observer gain vector determination, 181–182 open-loop observer, 179–180 optimal regulator, 208–215 pointer balancing, 190–191, 191 pole locations, 182–186 proof, 181 reduced-order observer, 191–195, 193 regulator, optimal, 208–215 rigid mode, 226 robust control problem, 223–224, 226–228 simple mass, 194, 194–195, 204–205, 212–213, 219–220 single-output system, 181–182 spring-mass damper, 198 state dynamics, 200, 208, 216 steady state, 208–215 stochastic forces, 198, 204–205, 208–215 structured parametric uncertainties, 222–224 theorem, 181 unmodeled dynamics, 224–228 variance, 197 white noise response, 195–200 Estimated states, robust control, 394–396 EVAD, see Electric ventricular assist device (EVAD) Exercise problems estimated states control, 228–233 robust control, 367–376, 402–403 state feedback control and optimization, 167–178 state space description, 73–77 Existence theorems, 156, see also Theorems

F Facts derivatives, 419, 421 full state feedback control, 317 gain margin, 257 H2 and H∞ norms, 274, 279–280 linear function, 421 matrix norms, 428 multiinput system, 257 optimization, scalar function with constraints, 444–445 phase margin, 257 positive definite matrix, 419 positive semidefinite matrix, 419 positive semidefinite quadratic function, 418 quadratic function, 419

robust control, sliding mode methods, 396 sensitivity, robustness, 240 single input, single output case, 139 state space description, 40 well-posedness and internal stability, 289–290 Feedback, see State feedback control and optimization Filter algebraic Riccati equation (FARE), see also Algebraic Riccati equation (ARE); Control algebraic Riccati equation (CARE); Riccati equation linear quadratic Gaussian control, 218, 264 observer dynamics, 203 First-order system, 132–134 Flexible mode, 226–228 Flexible/rigid modes system, 161–163 Flexible tetrahedral truss structure, 390–392, 447–449, see also Uncertain tetrahedral truss structure Forbidden unit circle, 255, 255 Fourier integrals, 422–425, see also Inverse Fourier transformations Frequency-shaped linear quadratic control fundamentals, 148–151 harmonic oscillator, 151–154 standard control, 152 standard linear quadratic control, 152 Friedland studies, 19 Fujita studies, 229, 375 Full state feedback, control H∞ control, 314–322 H2 control, 283, 283–286 state feedback control and optimization, 108–112 Full state feedback, linear system double integrator system, 384–385 fundamentals, 379–380 helicopter, 382–383 optimal sliding mode controller, 383–385 sliding hyperplane matrix, 380–383 Full state feedback, uncertain linear system flexible tetrahedral truss structure, 390–392 H∞ method, 386–394 impact, uncertainties, 387–388 Matlab programming, 392–394 sliding hyperplane matrix computation, 388–394 uncertain tetrahedral truss structure, 392–394

G Gain margin multiinput system, 257–260 single input system, LQR, 254–256


Index Gain matrix computation, multiinput system, 95–99 Gain vector computation, 80–85, 82 GCRD, see Greatest common right divisor (GCRD) Gopal studies, 92, 156, 160, 409–416 Gramian determinant, 414 Gramian observability and matrix, 19–21 Graphical interpretation, 360 Greatest common right divisor (GCRD), 47–49, 51–57 Greatest lower bound, 433 Green and Limebeer studies, 278 Guaranteed norm, 322–324 Gupta studies, 148 Gyro box, 101–108, 102, 108–112

H Hamiltonian autonomous system, 115–123 H∞ norm, 281 linear quadratic trajectory control, 144 linear time-invariant system, 154 LQR problem, root locus plot, 138 minimum-time control, 154, 167–168 optimality, 114 time-optimal control, 157 Harmonic oscillator, 151–154 H∞ control active suspension system, 319–320 cases, 335–336 controller design, 332–336 disturbance feedforward, 320–322 equivalence, 325, 328–332 facts, 317 full state feedback control, 314–322 guaranteed norm, 322–324 imaginary axis, 336–337 lemma, 325, 328, 330, 332, 334 output feedback control, 324–336 poles, imaginary axis, 336–337 proof, 317, 325–332, 334–335 simple mass, 317–319 stability, 328–332 state estimation, 322–324 theorems, 325 zeros, imaginary axis, 336–337 H2 control active suspension system, 285–286 double integrator system, 285, 287–288 full state feedback H2 control, 283, 283–286 output feedback H2 control, 286, 286–288 simple mass, 287–288

461 Helicopter, 382–383 High cost of control, 140 H∞ method, 386–394 Homogeneous equation, 12–13 Hughes studies, 146

I Imaginary axis, 336–337 Impact, uncertainties, 387–388 Independent vectors, 65 Inequality for robustness test, 237–238 Infinimum, 432–433 Infinite final time, 132–137 Inhomogeneous equation, 13–14 Inman, Schultz and, studies, 101–108 Innovation process, 203 Input loop transfer matrix, 288 Input return difference matrix, 288 Inputs and constraints, 114–115 Input shaping, 163–168 Integral action, 91–95 Integral output feedback, 94–95 Intensity matrix Kalman filter, 200 linear quadratic Gaussian control, 216 optimal sliding mode Gaussian control, 397 stochastic optimal regulator, 208 Internal stability, 288–292 Inverse Fourier transformations, 422, 439, see also Fourier integrals Inverse Laplace transformations, see also Laplace transformations H2 and H∞ norms, 269, 278 inhomogeneous equation, 13–14 state space equations, 12 Irreducible matrix-fraction description, 71

K Kailath studies balancing a pointer, 82 closed-loop observer, 180–181 combined observer-controller, 189 controller form realization, 59, 65 degree of a square polynomial matrix, 47, 49 duality, 10 greatest common right divisor, 51 infinite final time, 132 integral action, 93, 96–97 low cost of control, 142 LQR problem, root locus plot, 139 minimal order, state space realization, 58


462


multiinput/multioutput systems, 38–40 nonzero and constant output, 90 open-loop observer, 179 reduced-order observer, 191, 193 row and column operations, 50 simultaneous controllability and observability, 32 Smith form, 66 Smith-McMillan form, 67 stability analysis, 72 state feedback gain matrix, 100 state feedback gain vector, 81 state space realization nonuniqueness, 11 zeros of the closed-loop transfer function, 86 Kalman and Bucy studies, 182 Kalman filter double integrator system, 204–205 linear quadratic Gaussian control, 217 linear quadratic regulator feedback loops, 260–263, 261, 263–264 loop transfer recovery, 265 Matlab programming, 205–208 measurement equation, 200–201 observer dynamics, 201–202 observer pole locations, 182 optimal sliding mode Gaussian control, 397 simple mass, 204–205 state dynamics, 200 stochastic disturbances, 204–205 three-disk system, 399 Kao and Sinha studies, 394 Kao studies, 447 Kar studies, 370 Kirchoff’s law, 128 Kitada studies, 372 Kuo studies, 4, 128, 140, 163 Kwakernaak and Sivan studies, 132, 138, 266, 288, 440

L Lack of guaranteed robustness, 263–265 Lagrange multiplier, 202, 210 Laminar pipe flow, 26 Laplace transformations, see also Inverse Laplace transformations bicycle dynamics, 87 inhomogeneous equation, 13 Kalman filter, 260 noncolocated sensor and actuator, 143 robustness, 252 scalar signal, 422–423 SISO system, 3–4, 11 state space equations, 12

Law of conservation of mass, 26 Least upper bound, 433 Lemmas constant input disturbances, 92 controller design, 334 equivalence of stability systems, 328, 330, 332 H2 and H∞ norms, 273, 275–278 matrix inversion lemma, 415–416 MIMO system, 38 observer dynamics, 202–203 output feedback H∞ control, 325 stochastic optimal regulator, 210 LFT, see Linear fractional transformation (LFT) Limebeer, Green and, studies, 278 Linear algebraic equations, system, 409–414 Linear continuous-time system, white noise response fundamentals, 195–196 Matlab programming, 198–200 mean, 196–197 spring-mass damper, 198 stochastic forces, 198 variance, 197 Linear fractional transformation (LFT) structured uncertainties, robust control, 304–308, 305–306, 308 uncertainties impact, 388 Linear independence, vectors, 409, 414 Linear quadratic Gaussian (LQG) control closed-loop system dynamics, 218–219 double integrator, 268–269 double integrator system, 219–220 fundamentals, 215 lack of guaranteed robustness, 263–265 loop transfer recovery, 265–269 Matlab programming, 220–221, 313–314 measurement equation, 216 objective function, 216–219 proof, 266–268 simple mass, 219–220, 268–269 state dynamics, 216 theorems, 266 Linear quadratic maximization problem, 275–278 Linear quadratic regulator (LQR) active suspension, 134–137 closed-loop optimal control, 126–127 cost of control, 140–142 cross term, 127–132 DC motor, 128–132 energy control, minimum, 128–132 fact, 139 first-order system, 132–134 fundamentals, 123–137 high cost of control, 140 infinite final time, 132–137


Index low cost of control, 140–142 minimum energy control, 128–132 noncolocated sensor/actuator, 142–143 objective function, 127–132 open-loop optimal control, 124–126 optimal linear state feedback, 134–137 sensor, noncolocated, 142–143 SISO system, 137–143 symmetric root locus, 139–142 Linear quadratic regulator (LQR), feedback loops fundamentals, 252, 252–254 gain margin, 254–260 Kalman filter, 260–263, 261, 263–264 linear quadratic regulator, 252–263 multiinput system, 257–260 phase margin, 254–260 proof, 253–254, 258 simple mass, 255–256, 262–263 single-input system, 254–256 special case, 259 Linear quadratic trajectory control fundamentals, 143–146, 147 sun tracking solar concentrations, 146–148 Linear state feedback, optimal, 134–137 Linear systems, full state feedback double integrator system, 384–385 fundamentals, 379–380 helicopter, 382–383 optimal sliding mode controller, 383–385 sliding hyperplane matrix, 380–383 Linear systems, robust control, 394–396 Linear time-invariant system, minimum-time control cases, 158–159 double integrator system, 156–160 existence theorems, 156 flexible/rigid modes system, 161–163 fundamentals, 154–155 harmonic oscillator, 151–154 input shaping, 163–168 linear time-invariant system, 154–168 normality, linear systems, 155–156 optimality, 166–168 rigid body time-optimal control, 156–160 rigid/flexible modes system, 161–163 theorems, 156 time-optimal control, 156–160 time optimality, 166–168 uniqueness theorems, 156 Loop shaping Bode’s sensitivity integrals, 345–351 constraints, 345–351 H∞ control, 337–344 Matlab programming, 343–344, 348–351 performance/robustness trade-off, 337–344

463 single-input single-output system, 345–351 three-disk system, 340–343 X-29 aircraft, 347–351 Loop transfer recovery (LTR) double integrator, 268–269 lack of guaranteed robustness, 263–265 loop transfer recovery, 265–269 proof, 266–268 simple mass, 268–269 theorems, 266 Low cost of control, 140–142 Lowell and McKell studies, 86 Lower linear fractional transformation, 305–306 LQG, see Linear quadratic Gaussian (LQG) control LTR, see Loop transfer recovery (LTR) Lublin and Athans studies, 390 Lyapunov equation H2 and H∞ norms, 279–280 H∞ control, 316–317 observer dynamics, 203 stability analysis, 72 stochastic optimal regulator, 210 white noise response, linear continuous-time system, 197–198

M Maciejowski studies, 236, 432 Matlab commands lyap, 281 normh2, 281 normhinf, 283 null, 382 sigma, 237 Matlab programming approximate Gaussian white noise generation, 439, 440 Bode’s sensitivity integrals, 343–344 electronics navigation, 108–112 full state feedback control, 108–112 Gaussian white noise generation, 439, 440 gyro box, 108–112 Kalman filter, 205–208 linear quadratic control, 313–314 linear quadratic Gaussian control, 220–221, 247–251 linear system, white noise response, 198–200 loop shaping, 343–344 sensitivity, robustness, 247–251 spring-mass system, 247–251 stochastic optimal regulator, 214–215 structured singular values analysis, 364–366


464


structured uncertainties, robust control, 313–314 tetrahedral truss structure, uncertain, 392–394 uncertain tetrahedral truss structure, 392–394 white noise generation, 439, 440 white noise response, linear system, 198–200 X-29 aircraft, 348–351 Matlab routines aresolv, 392 augtf, 347 fitd, 367 hinf, 342 hinfopt, 347, 363, 367 lyap, 198 ODE23, 133 ODE45, 133 randn, 439 ssv, 367 Matrices norms, 427–429 range, 409 similar, 11 singular values, 429 Matrix-fraction description (MFD) controller form realization, 59–66 degree, square polynomial matrix, 47–49 fundamentals, 46–47 greatest common right divisor, 47–49 gyro box, 103 minimal order, state space realization, 57–59 similarity transformation, 64–66, 66 square polynomial matrix, 47–49 state space, 59–64, 71 transfer function matrix, 57–59 Matrix inversion lemma controller design, 334 fundamentals, 415–416 loop transfer recovery, 266–267 Maximization, optimality conditions, 115 Maximum distance, 117–123 Maximum singular values, 431–432 McKell, Lowell and, studies, 86 Mean, 196–197, 441 Measurement equation Kalman filter, 200–201 linear quadratic Gaussian control, 216 optimal sliding mode Gaussian control, 396–397 Meditch studies, 439 Methods controllability, 24–25, 45–46 H2 and H∞ norms, 278–280 observability, 16–17, 32–36, 45–46 SISO system, 5–9

state space realization, 5–9, 16–17, 24–25, 32–36 transfer function matrix, 40–44 MFD, see Matrix-fraction description (MFD) Miller, Sinha and, studies, 384, 395–397, 400 MIMO, see Multiinput/multioutput systems (MIMO) Minimal order, state space description, 57–59 Minimization, optimality conditions, 114 Minimum energy control, 128–132 Minimum singular values, 431–432 Minimum-time control, linear time-invariant system cases, 158–159 double integrator system, 156–160 existence theorems, 156 flexible/rigid modes system, 161–163 fundamentals, 154–155 harmonic oscillator, 151–154 input shaping, 163–168 linear time-invariant system, 154–168 normality, linear systems, 155–156 optimality, 166–168 rigid body time-optimal control, 156–160 rigid/flexible modes system, 161–163 theorems, 156 time-optimal control, 156–160 time optimality, 166–168 uniqueness theorems, 156 Modal compatibility, 82 Modeling errors impact benchmark robust control, 223–224, 226–228 flexible mode, 226–228 rigid mode, 226 structured parametric uncertainties, 222–224 unmodeled dynamics, 224–228 Model properties, state space description, 10–11 Moore, Anderson and, studies, 127, 132, 203, 239–242, 257 Morari, Braatz and, studies, 360 Multiinput/multioutput systems (MIMO) analog computer simulation diagram, 43 Bode’s sensitivity integrals, 346 column-reduced matrix, 38 definitions, 37–38, 44–45 frequency-shaped LQ control, 150 fundamentals, 36–37 lemma, 38 methods, 45–46 poles and zeros, 66–71 Smith form, 66–67 Smith-McMillan form, 67–71 transfer function matrix, 66–71


Index Multiinput system linear quadratic regulator feedback loops, 257–260 state feedback gain matrix computation, 95–99 Multiple output system, closed-loop observer, 182 Multiplicative uncertainty sensitivity, robustness, 246–247 unstructured uncertainty, 299–301, 300, 302–304 Multivariable Parseval relation, 425

N Necessity, small gain theorem, 293–295 Newton’s second law of motion bicycle dynamics, 86 DC motor, minimum energy control, 129 gyro box, 101 pointer balancing, 83 No constraint on input, 114 Noncolocated sensor/actuator, 142–143 Nonsingular polynomial matrix, 40 Nonuniqueness, 10–11, 49 Nonzero output, 89–91 Normality, linear systems, 155–156 Notation, SISO system state space realizations, 9 Null space (nullity), 410–414 Nyquist criterion gain and phase margins, 258 simple mass control, 256, 256–257 unstructured uncertainties, 244–245

O Objective function linear quadratic Gaussian control, 216–219 linear quadratic regulator, 127–132 stochastic optimal regulator, 209–211 Observability matrix, 16 MIMO system, 44–46 simultaneous, 31–36 SISO system, 14–19 state space realizations, 5 Observation spillover, 225 Observer-based control, 222–228 Observer dynamics, 201–202 Observer gain matrix determination, 182 Observer gain vector determination, 181–182 Observer poles, 182–186 Open-loop observer, 179–180 Open-loop optimal control, 124–126

465 Open-loop stability, 84 Optimal control theory autonomous system, 115–123 conditions, optimality, 113–115 constraints, 114–115, 117–120 fundamentals, 112–113 Hamiltonian, 115–123 inputs and constraints, 114–115 maximization, 115 maximum distance, 117–123 minimization, 114 no constraint on input, 114 special case, 116–117 time, 117–123 travel, velocity constraints, 117–123 velocity constraints, 117–123 zeros, 120–123 Optimality, minimum-time control, 166–168 Optimal linear state feedback, 134–137 Optimal regulator, 208–215 Optimal sliding mode Gaussian (OSG) control measurement equation, 396–397 proof, 398–399 state dynamics, 396–397 theorems, 397–398 three-disk system, 399–400 Optimal sliding mode (OS) controller, 383–385 Optimization, scalar function with constraints, 443–445 OS, see Optimal sliding mode (OS) controller OSG, see Optimal sliding mode Gaussian (OSG) control Output feedback control, 286–288, 324–336 Output loop transfer matrix, 288 Output return difference matrix, 288

P Pai and Sinha studies, 388 Pao, Singhose and, studies, 163 Pao and Singhose studies, 163 Papoulis studies, 435 Parseval’s theorem frequency-shaped LQ control, 148–149 fundamentals, 422–425 H2 and H∞ norms, 269–270, 272, 275, 278 Penn State electric ventricular assist device (EVAD), 182–184, 183 Performance, controller, 354–366 Performance/robustness trade-off, 337–344 Periodic structure, SISO system, 27–28 Phase margin multiinput system, 257–260 single input system, 254–256


466


Plant noise vector, 144, see also Process noise Plant transfer matrix, 299, 301 Pointer balancing combined observer-controller, 190–191, 191 state feedback gain vector computation, 82–85 Poles closed-loop transfer function, 80–85, 82 imaginary axis, 336–337 irreducible matrix-fraction description, 71 locations, closed-loop observer, 182–186 Smith-McMillan form, 66–71 Polynomial vectors, 38 Position output integral output feedback, 94 spring-mass-damper system, 17–18 spring-mass system, 90–91 Positive definite matrix, 419 Positive definite quadratic function, 418 Positive semidefinite matrix, 419 Positive semidefinite quadratic function, 418 Power spectrum/spectral density (PSD), 438–439, 441 Process noise, see also Plant noise vector Kalman filter, 200 linear quadratic Gaussian control, 216 optimal sliding mode Gaussian control, 396 stochastic optimal regulator, 208 Proofs closed-loop observer, 181 constant input disturbances, 92–93 controller design, 334–335 derivatives, 420–422 equivalence of stability systems, 328–332 full state feedback control, 317 H2 and H∞ norms, 274–275, 279–280 linear quadratic regulator feedback loops, 253–254, 258 loop transfer recovery, 266–268 matrix inversion lemma, 415–416 matrix norms, 428–429 multivariable Parseval relation, 425 optimal sliding mode Gaussian control, 398–399 output feedback H∞ control, 325–328 scalar Parseval relation, 424 sensitivity, robustness, 240 SISO system, controllability, 22–24 well-posedness and internal stability, 289–290, 293–295 Properties, controller, 351–353

Q Quadratic functions, 417–419 Quarter car model, 134, 134

R Range of a matrix, 409 Range space, 409 Rational transfer function matrix, 37 Ray studies, 112–113, 116, 126 Reaching condition, 377 Realization, state space description, 5–11, 7–8, 16–17 Reddy studies, 432–433 Redheffer studies, 332 Reduced-order observer, 191–195, 193 References, 405–408 Regular point, 443–444 Regulator, optimal, 208–215 Repeated blocks, 351 Residual dynamics, 225 Residual vibratory mode, 302–304 Return difference equality, 252–254 Riccati equation, see also Algebraic Riccati equation (ARE); Control algebraic Riccati equation (CARE); Filter algebraic Riccati equation (FARE) closed-loop optimal control, 126 controller design, 333 first-order system, 133 stochastic optimal regulator, 210 Riccati-like equation, 390 Rigid body mode, 341–342 Rigid body time-optimal control, 156–160 Rigid/flexible modes system, 161–163 Rigid mode, 226 Robust control active suspension system, 285–286, 319–320 additive uncertainty, 301, 301–304 benchmark robust control, 360–364 Bode’s sensitivity integrals, 345–351 bounded real lemma, 273, 275–278 cases, 295–297, 335–336 complementary sensitivity, 238–251 computations, 278–283 constraints, 345–351 controller, structured singular values analysis, 351–367 controller design, 332–336 definitions, 238–242, 269–270, 273, 289 disturbance feedforward, 320–322 D-K iteration, 366–367 double integrator, 268–269, 285, 287–288


Index equivalence, 325, 328–332 exercise problems, 367–376 facts, 240, 257, 274, 279–280, 290, 317 feedback loop, 252, 252–254 full state feedback control, 283, 283–286, 314–322 gain margin, 254–260 graphical interpretation, 360 guaranteed norm, 322–324 H2 and H∞ norms, 269–283 H∞ control, 314–337, 343–344 H2 control, 283–288 imaginary axis, 336–337 inequality, test, 237–238 internal stability, 288–292 Kalman filter, 260–263, 261, 263–264 lack of guaranteed robustness, 263–265 lemma, 273, 275–278, 325, 328, 330, 332, 334 linear fractional transformation, 304–308, 305–306, 308 linear quadratic Gaussian control, 263–269, 313–314 linear quadratic maximization problem, 275–278 linear quadratic regulator, 252–263 loop shaping, 337–351 loop transfer recovery, 265–269 lower linear fractional transformation, 305–306 Matlab programming, 247–251, 313–314, 343–344, 348–351, 364–366 methods, 278–280 modeling errors impact, 223–224, 226–228 multiinput system, 257–260 multiplicative uncertainty, 246–247, 299–301, 300, 302–304 necessity, 293–295 output feedback control, 286, 286–288, 324–336 performance, 337–344, 354–366 phase margin, 254–260 poles, imaginary axis, 336–337 residual vibratory mode, 302–304 robustness analysis, 353–354 root mean square response-, 271–273 sensitivity, 238–251 significances, 270–278 simple mass, 255–256, 262–263, 268–269, 287–288, 317–319 single-input, single-output system, 345–351, 357–360 single-input system, 254–256 singular value analysis, 235–238 small gain theorem, 292–299, 313 special case, 259

467 spring-mass damper system, 236–237, 281, 283 spring-mass system, 247–251 stability, 328–332 state estimation, 322–324 state space, 291 state space model, 307–308, 308 structured parametric uncertainties, 308–314 structured singular values, 351–353 structured uncertainties, 242, 242–244, 304–314 sufficiency, 293 synthesis, 366–367 three-disk system, 340–343 transfer function, 291–292 transfer function matrix, 235–237 two-mass system, 311–313 uncertain stiffness, 311–313 unit impulse responses, 270–271 unstructured uncertainties, 244, 244–251, 299–304 upper linear fractional transformation, 306–307 well-posedness, 288–290 worst-case frequency response, 273–274 worst-case output or input, 274–275 X-29 aircraft, 347–351 zeros, imaginary axis, 336–337 Robust control, sliding mode methods double integrator system, 384–385 estimated states, 394–396 exercise problems, 402–403 facts, 396 flexible tetrahedral truss structure, 390–392 full state feedback, 379–394 fundamentals, 377–379, 378 helicopter, 382–383 H∞ method, 386–394 impact, uncertainties, 387–388 linear systems, 379–385, 394–396 Matlab programming, 392–394 measurement equation, 396–397 optimal slide mode Gaussian control, 396–400 optimal sliding mode controller, 383–385 proof, 398–399 sliding hyperplane matrix, 380–383, 388–394 state dynamics, 396–397 theorems, 397–398 three-disk system, 399–400 uncertain linear system, 386–394 uncertain tetrahedral truss structure, 392–394 Root locus plot, SISO case cost of control, 140–142 fact, 139 noncolocated sensor/actuator, 142–143


468


sensor, noncolocated, 142–143 symmetric root locus, 139–142 Root mean square response-, 271–273 Row operations, 50–51 Rugh studies, 50

S Safanov, Chiang and, studies, 336, 338, 366, 432 Safanov studies, 336 Sage and White studies, 143–145 Scalar Parseval relation, 424 Schmitz, Cannon and, studies, 232 Schultz and Inman studies, 101–108 Seering, Singer and, studies, 165 Sensitivity, robustness definitions, 240 facts, 240 fundamentals, 238–242 Matlab programming, 247–251 multiplicative uncertainty, 246–247 proof, 240 spring-mass system, 247–251 structured uncertainties, 242, 242–244 unstructured uncertainties, 244, 244–251 Sensor, noncolocated, 142–143 Seo and Chen studies, 423 Separation property, 188 Separation theorem, 217 Shaped input, 163–168 Shultz and Inman studies, 101–108 Significances, H2 and H∞ norms, 270–278 Similarity transforms and transformations controller form realization, 64–66, 65 state space description, 29–31 Similar matrices, nonuniqueness, 11 Simple mass full state feedback control, 317–319 H2 control, 287–288 Kalman filter, 204–205, 262–263 linear quadratic Gaussian control, 219–220 linear quadratic regulator feedback loops, 255–256 loop transfer recovery, 268–269 reduced-order observer, 194, 194–195 stochastic optimal regulator, 212–213 Simultaneous controllability and observability, 31–36 Singer and Seering studies, 165 Singhose, Pao and, studies, 163 Singhose and Pao studies, 163 Single-input/single-output (SISO) system Bode’s sensitivity integrals, 345–351 case, 139–142

cost of control, 140–142 duality, 10 fact, 139 methods, 5–9 multiplication uncertainty, 246–247 noncolocated sensor/actuator, 142–143 nonuniqueness, 10–11 notation, 9 observability, 14–19, 18 robustness analysis, 357–360 sensor, noncolocated, 142–143 SISO system, 137–143 state space model properties, 10–11 state space realizations, 5–12, 7–8 stationary stochastic process response, 440, 440 symmetric root locus, 139–142 transfer function, 3–5, 11–12 Single-input system, linear quadratic regulator feedback loops, 254–256 Single-input system, state variable feedback bicycle dynamics, 86–89 closed-loop transfer function, 80–89 constant input disturbances, 91–95 constant output, 89–91 controllability, 84 fundamentals, 79–80 gain vector computation, 80–85, 82 integral action, 91–95 integral output feedback, 94–95 lemma, 92 open-loop stability, 84 pointer balancing, 82–85 poles, closed-loop transfer function, 80–85, 82 position output, 90–91, 94 proof, 92–93 single input system, 79–95, 80 spring-mass system, 90–91 velocity output, 91, 95 zeros, 85–89, 86 Single-output system, closed-loop observer, 181–182 Singular value decomposition (SVD), 429–431 Singular values analysis, 235–238 matrix, 429 properties, 432 Sinha, Kao and, studies, 394 Sinha, Pai and, studies, 388 Sinha and Miller studies, 384, 395–397, 400 SISO, see Single-input/single-output (SISO) system Sivan, Kwakernaak and, studies, 132, 138, 266, 288, 440 Sliding hyperplane matrix, 380–383, 388–394


Index Slotine, Asada and, studies, 377, 387 Small gain theorem additive uncertainty, 302 multiplicative uncertainty, 300 robust control, 292–299 structured uncertainties, robust control, 313 Smith form, 66–67 Smith-McMillan form, 67–71 Space description, see State space Space shuttle reentry dynamics, 451–455 Special case Hamiltonian, autonomous system, 116–117 linear quadratic regulator feedback loops, 259 matrix inversion lemma, 416 Spectral radius, 353 Spillover, control and observation, 225 Spring-mass damper, stochastic forces, 198 Spring-mass damper system H2 and H∞ norms, 281, 283 integral output feedback, 94 observability matrix, 17–19 singular value analysis, 236–237 Spring-mass system nonzero and constant output, control, 90–91 sensitivity, robustness, 247–251 Square polynomial matrix, 47–49 Stability internal, 288–292 open-loop stability, 84 output feedback H∞ control, 328–332 state space, 71–72 Staircase input, 164, 164 Standard linear quadratic control, 152 State dynamics Kalman filter, 200 linear quadratic Gaussian control, 216 optimal sliding mode Gaussian control, 396–397 stochastic optimal regulator, 208 State error dynamics, 195 State estimation, 322–324 State feedback control and optimization active suspension, 134–137 actuator, noncolocated, 142–143 autonomous system, 115–123 bicycle dynamics, 86–89 cases, 158–159 closed-loop optimal control, 126–127 closed-loop transfer function, 80–89 conditions, optimality, 113–115 constant input disturbances, 91–95 constant output, 89–91 constraints, 114–115, 117–120 controllability, 84 cost of control, 140–142

469 cross term, 127–132 DC motor, 128–132 double integrator system, 156–160 eigenvalues and eigenvectors, 99–112, 102 electronics navigation, 101–112 energy control, minimum, 128–132 exercise problems, 168–178 existence theorems, 156 fact, 139 first-order system, 132–134 flexible/rigid modes system, 161–163 frequency-shaped linear quadratic control, 148–154 full state feedback control, 108–112 gain matrix computation, multiinput system, 95–112 gain vector computation, 80–85, 82 gyro box, 101–112 Hamiltonian, 115–123 harmonic oscillator, 151–154 high cost of control, 140 infinite final time, 132–137 inputs and constraints, 114–115 input shaping, 163–168 integral action, 91–95 integral output feedback, 94–95 lemma, 92 linear quadratic regulator, 123–143 linear quadratic trajectory control, 143–148, 147 linear time-invariant system, 154–168 low cost of control, 140–142 Matlab programming, 108–112 maximization, 115 maximum distance, 117–123 minimization, 114 minimum energy control, 128–132 minimum-time control, 154–168 multiinput system, gain matrix computation, 95–112 no constraint on input, 114 noncolocated sensor/actuator, 142–143 nonzero output, 89–91 normality, linear systems, 155–156 objective function, 127–132 open-loop optimal control, 124–126 open-loop stability, 84 optimal control theory, 112–123 optimality, 166–168 optimal linear state feedback, 134–137 pointer balancing, 82–85 poles, closed-loop transfer function, 80–85, 82 position output, 90–91, 94 proof, 92–93 rigid body time-optimal control, 156–160


470


rigid/flexible modes system, 161–163 root locus plot, 137–143 sensor, noncolocated, 142–143 single input system, 79–95, 80 SISO system, 137–143 special case, 116–117 spring-mass system, 90–91 standard linear quadratic control, 152 sun tracking solar concentrations, 146–148 symmetric root locus, 139–142 theorems, 156 time, 117–123 time-invariant system, 154–168 time-optimal control, 156–160 time optimality, 166–168 travel, velocity constraints, 117–123 uniqueness theorems, 156 variable, 79–95 velocity constraints, 117–123 velocity output, 91, 95 zeros, 85–89, 86, 120–123 State space canonical form, 30–31 column operations, 50–51 column-reduced matrix, 38 controllability, 19–28, 30–36, 44–46 controller form realization, 59–66 definitions, 37–38, 44–45, 49 degree of a square polynomial matrix, 47–49 diagonal form, 29–30 duality, 10 equations solutions, 12–14 exercise problems, 73–77 facts, 40 greatest common right divisor, 47–49, 51–57 homogeneous equation, 12–13 inhomogeneous equation, 13–14 internal stability, 291 irreducible matrix-fraction description, 71 lemmas, 38 Lyapunov equation, 72 matrix-fraction description, 46–66, 71 methods, 5–9, 16–17, 24–25, 32–36, 40–46 minimal order, 57–59 models, 10–11, 307–308 multiinput/multioutput systems, 36–40, 44–46, 66–71 nonuniqueness, 10–11, 49 notation, 9 observability, 14–19, 31–36, 44–46 periodic structure, 27–28 poles, 66–71 position output, 17–18 proofs, 22–24 properties, 10–11

realization, 5–11, 7–8, 16–17 row operations, 50–51 similarity transforms and transformations, 29–31, 64–66 simultaneous controllability and observability, 31–36 single input/single output system, 3–12, 14–28 Smith form, 66–67 Smith-McMillan form, 67–71 space shuttle reentry dynamics, 451–455 square polynomial matrix, 47–49 stability analysis, 71–72 structured uncertainties, 307–308, 308 theorems, 22, 32 transfer function, 3–5, 4, 11–12 transfer function matrix, 40–44, 57–59, 66–71 velocity output, 18–19 zeros, 66–71 State transition matrix, 13 State variable feedback, single-input system bicycle dynamics, 86–89 closed-loop transfer function, 80–89 constant input disturbances, 91–95 constant output, 89–91 controllability, 84 fundamentals, 79–80 gain vector computation, 80–85, 82 integral action, 91–95 integral output feedback, 94–95 lemma, 92 nonzero output, 89–91 open-loop stability, 84 pointer balancing, 82–85 poles, closed-loop transfer function, 80–85, 82 position output, 90–91, 94 proof, 92–93 single input system, 79–95, 80 spring-mass system, 90–91 velocity output, 91, 95 zeros, 85–89, 86 State variables, 6 Stationary stochastic processes, 440–441 Steady state covariance matrix, 209 error covariance, 202 estimated states control, 208–215 Stein, Doyle and, studies, 244, 268, 382 Stein studies Bode’s sensitivity integrals, 345 controller design, 335 full state feedback control, 315–316 output feedback control, 324–325 root mean square response, 272 system stability, 328, 332


Index worst-case output/input, 274 X-29 aircraft, 347 Stochastic forces and processes autocorrelation function, 436–437 autocovariance function, 437 crosscorrelation function, 437 crossvariance function, 437 fundamentals, 435 Kalman filter, 204–205 optimal regulator, steady state, 208–215 power spectrum/spectral density, 438–439 spring-mass damper, 198 stationary, 435–438, 440–441 uncorrelated processes, 438 vectors, 440–441 white noise, 439 Strang studies, 380, 384, 409–416, 418–419 Strictly proper transforms, 269 Structured parametric error, 222 Structured parametric uncertainties modeling errors impact, 222–224 structured uncertainties, robust control, 308–314 Structured uncertainties sensitivity, robustness, 242, 242–244 structured uncertainties, robust control, 304–314 Structured uncertainties, robust control linear fractional transformation, 304–308, 305–306, 308 linear quadratic Gaussian control, 313–314 lower linear fractional transformation, 305–306 Matlab programming, 313–314 small gain theorem, 313 state space model, 307–308, 308 structured parametric uncertainties, 308–314 structured uncertainties, 304–314 two-mass system, 311–313 uncertain stiffness, 311–313 upper linear fractional transformation, 306–307 Strunce studies, 447 Sufficiency, small gain theorem, 293 Sun tracking solar concentrations, 146–148 Supremum, 432–433 SVD, see Singular value decomposition (SVD) Symmetric root locus, 139–142 Synthesis, controller, 366–367 System, linear algebraic equations, 409–414

T Tank-pipe system, 25

471 Taylor series expansion, 116 Theorems closed-loop observer, 181 H2 and H∞ norms, 281–283 linear quadratic Gaussian control, 217 loop transfer recovery, 266 minimum-time control, 156 null space (nullity), 410–414 optimal sliding mode Gaussian control, 397–398 output feedback H∞ control, 325 robustness analysis, 353–354 separation, 217 simultaneous controllability and observability, 32 SISO system, controllability, 22 Thompson studies, 134, 137 Three-disk system, 340–343, 399–400 Time, 117–123, 132–137 Time-optimal control, 156–160 Time optimality, 166–168 Toeplitz matrix, upper triangular, 81 TPBVP, see Two-point boundary value problem (TPBVP) Transfer function and transfer function matrix internal stability, 291–292 MFD, minimal order, 57–59 poles and zeros, MIMO, 66–71 singular value analysis, 235–237 state space, 3–5, 4, 11–12, 40–44 Travel, velocity constraints, 117–123 Triangular Toeplitz matrix, upper, 81 Tsach studies, 182–183 Two-degree-of-freedom system, 142, 142 Two-mass system, 311–313 Two-point boundary value problem (TPBVP), 125–126

U Uncertain linear system, full state feedback flexible tetrahedral truss structure, 390–392 H∞ method, 386–394 impact, uncertainties, 387–388 Matlab programming, 392–394 sliding hyperplane matrix computation, 388–394 uncertain tetrahedral truss structure, 392–394 Uncertain stiffness, 311–313 Uncertain tetrahedral truss structure, 392–394, see also Flexible tetrahedral truss structure Uncorrelated stochastic processes, 438 Unimodularity, 49, 51


472


Uniqueness theorems, 156 Unit impulse responses, 270–271 Unmodeled dynamics, 224–228 Unstructured uncertainty additive uncertainty, 301, 301–304 fundamentals, 299 multiplicative uncertainty, 299–301, 300, 302–304 residual vibratory mode, 302–304 sensitivity, robustness, 244, 244–251 Upper linear fractional transformation, 306–307 Upper triangular Toeplitz matrix, 81 Utkin and Yang studies, 380–381 Utkin studies, 377

V Variables, state space realizations, 6 Variance, 197, 441 Vectors independence, 65 linear independence, 409, 414 norms, 427 polynomial, multiinput/multioutput systems, 38 signal, 424–425 singular value decomposition, 431–432 stationary stochastic processes, 440–441 Velocity constraints, 117–123 Velocity output integral output feedback, 95 spring-mass damper system, 18–19 spring-mass system, 91 Vidyasagar studies, 274, 427–428

W Well-posedness, 288–290 White, Sage and, studies, 143–145 White noise response, linear continuous-time system fundamentals, 195–196, 439 Matlab programming, 198–200 mean, 196–197 spring-mass damper, 198

stochastic forces, 198 variance, 197 Wiberg studies, 8 Wie and Bernstein studies, 223–224 Worst-case frequency response, 273–274 Worst-case output or input, 274–275

X X-29 aircraft, 347–351

Y Yang, Utkin and, studies, 380–381 Yang studies, 176

Z Zero-mean Gaussian white noise, 198 Zero-mean white noise inputs, 273 Zero-mean white noise vector, 196, see also White noise response, linear continuoustime system Zeros closed-loop transfer function, 85–89, 86 imaginary axis, 336–337 irreducible matrix-fraction description, 71 Smith-McMillan form, 66–71 state feedback control and optimization, 120–123 Zhou studies Bode’s sensitivity integrals, 345–346 detectability, 181 D-K iteration, 366 H∞ norm, 282 linear fractional transformation, 304 robustness analysis, 354 scalar function optimization, 443–445 singular value decomposition, 429 singular values, 351, 432 sliding hyperplane matrix, 390 small gain theorem, 293 structured parametric uncertainties, 308–309 structured singular values, 351 well-posedness, internal stability, 288–292

Linear systems: optimal and robust control

Linear Systems: Optimal and Robust Control

Robust and optimal control

Robust and optimal control

Linear Optimal Control

Robust control design. An optimal control approach

Optimal Control Systems

Optimal Adaptive Control Systems

Robust Control in Power Systems

Mathematical methods in robust control of linear stochastic systems

Mathematical Methods in Robust Control of Linear Stochastic Systems

Robust Control and Filtering of Singular Systems

Almost Optimal Adaptive Control Systems

Optimal Control

Infinite Dimensional Linear Control Systems: The Time Optimal and Norm Optimal Problems

Control Theory for Linear Systems

Optimal control and estimation

Optimal control

Control of Continuous Linear Systems

Optimization and Optimal Control

Linear state-space control systems

sol linear robust

Linear State-Space Control Systems

Optimal Sampled-Data Control Systems (Communications and Control Engineering)

Robust Control Of Time Delay Systems

Robust Adaptive Control of Uncertain Nonlinear Systems

Analysis and Control of Linear Systems

Linear Systems Control: Deterministic and Stochastic Methods

Linear Systems Control: Deterministic and Stochastic Methods

Linear Systems and Control: An Operator Perspective

Analysis and control of linear systems

Linear systems: optimal and robust control