Scribd
Upload
308 views76 pages

Robust Process Control Engineering

good book for control engineering

Uploaded by

pankaj
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
308 views76 pages

Robust Process Control Engineering

good book for control engineering

Uploaded by

pankaj
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
ROBUST PROCESS CONTROL 4¢¢< Manfred Morari Chemical Engineering California Institute of Technology Evanghelos Zafiriou Department of Chemical and Nuclear Engineering and Systems Research Center University of Maryland Prentice Hall Englewood Cliffs, New Jersey 07632 Library of Congress Cataloging-in- Publication Data ‘Morar, Mant. ‘Robust process contol / Mantied Morar, Evanghelos Zafirion om Bibliography: p. Includes index ISBN 0-13-789159-0 Chemical proce contrl. 1. Zafiriou, Bvangheos. ML Title TP1S5.75.MOT_ 1989 5.22028 660.2'81--de19 cP ‘Manufacturing buyer: Mary Ann Gloriande BAZ © 1989 by Prentice-Hall, Inc. ZZ8e A Division of Simon & Schuster Englewood Cliffs, New Jersey 07632 ‘The fonts in this book are Computer Modern Roman, set using Leslie Lam- port's IX,X document preparation facility, with the help of Jan Owen. Some of the figures were prepared by Glenn C., Smith on an Apple Macintosh with the program MacDraw. ‘TEX is a trademark of the American Mathematical Society. Macintosh and MacDraw are trademarks of Apple Computer, Inc. Alll rights reserved. No part of this book may be reproduced, in any form or by any means, without permission in writing from the publisher. Printed in the United States of America w98765432 ISBN 0-13-782153-0 Prentice-Hall International (UK) Limited, London Prentice-Hall of Australia Pty. Limited, Sydney Prentice-Hall Canada Inc., Toronto Prentice-Hall Hispanoamericana, 8.A., Mexico Prentice-Hall of India Private Limited, New Delhi Prentice-Hall of Japan, Inc., Tokyo Simon & Schuster Asia Pte. Ltd., Singapore Editora Prentice-Hall do Brasil, Ltda., Rio de Janeiro In memoriam CONSTANTIN G. ECONOMOU 1958 - 1986 Contents PREFACE xv NOMENCLATURE xix 1 INTRODUCTION 1 1.1 The Evolution of Control Theory 1 1.2 Controller Parametrization: The IMC Structure . 3 13. Robustness =e és 4 14 Scope of Book... .... 0... 4 1.5 Some Hints for the Reader 5 Part I: CONTINUOUS SINGLE-INPUT SINGLE-OUTPUT SYSTEMS a 2 FUNDAMENTALS OF SISO FEEDBACK CONTROL un 21 Definitions... .. . 2a ee u 22 Formulation of Control Problem See aes Ai 221 Process Model : Rohan eed 222 Model Uncertainty Description ve 6 223 Input Specification ........ 0000... 19 224 Control Objectives... BSA 2s 2 28 Internal Stability 2.00.0. se esse eee 2 2.4 Nominal Performance aes 2.4.1 Sensitivity and Complementary Sensitivity Function |... 24 24.2 Two-Degroe-of-Freedom Controller 1 35 24.3 Asymptotic Properties of the Closed-Loop Response Ge tem Type)... . 27 2.4.4 Linear Quadratic (H,-) Optimal Control 28 24.5 Hoy-Optimal Control... 1. sss ss 1. 39 25 Robust Stability : Ee ts oe SL 26 Robust Performance ... . nee rege ener 2.6.1 Hy Performance Objective... 0... 0s. + see 26.2 Hx Performance Objective 35 CONTENTS 27 Summary ..... 2.8 References... . ‘THE SISO IMC STRUCTURE FOR STABLE SYSTEMS 3.1 IMC Structure ante ak 32 Stability Conditions for IMC... +... « 3.2.1 Internal Stability 322 Relationship with “Clamic Feedback”... 1. 3.2.3 Implementation: Classic Feedback Versus IMC 33 Performance of IMC .........-00-0-eseeeee 3.3.1 Sensitivity and Complementary Sensitivity Function... 3.3.2. Two-Degree-of-Freedom Controller . . 3.3.3 Asymptotic Properties of Closed- Loon Response (System ype) a ost arate 3.34 The Concept of “Perfect Control 3.4 Outline of the IMC Design Procedure ..... « 3.4.1 Basic Design Philosophy : 34.2 Two-Degtee-of Freedom Design . . . « 3.43 Design in the Presence of Measurement Device Dynamics SS Sunamary » setlatayeat wie Paketaete er dt-w i a 3.6 References . . . SISO IMC DESIGN FOR STABLE SYSTEMS 4.1 Nominal Performance ...........+ 4.1.1 Hy-Optimal Control Per oe nplte 41.2 Design for Specific Input Forms»... 1.0... 4.1.3. Minimum Error Norm for Step Inputs 4.2 The IMC Filter . . 4.2.1 Filter Form 422 A Qualitative Interpretation of the Function of the Filter 4.3. Robust Stability f 4.3.1 Norm-Bounded Uncertainty Regions 4.3.2 General Uncertainty Regions . 4.4 Robust Performance , 44.1 Norm-Bounded Uncertainty Regions 44.2 General Uncertainty Regions . 4.5 Summary of IMC Design Procedure : 4.6 Application: IMC Design for a First-Order Deadtime System 4.6.1 Deadtime Uncertainty . 4.6.2 Three-Parameter Uncertainty 36 38 39 39 4 4 2 44 45 45 46 46 a7 50 50 52 54 55 55 37 CONTENTS: vi 47 References... + 84 5 SISO IMC DESIGN FOR UNSTABLE SYSTEMS BA 85 Parametrization of All Stabilizing Controllers . ody oa 8 5.1.1 Conditions for Internal Stability ©... 2... 00-0 85 5.1.2 Controller parametrization... 6... « 87 88 88 5.2. Nominal Performance é eS wees 5.2.1 HyOptimal Controller... 0.2.00 eee e eee ee 5.2.2 Design for Common Input Forms . - 9 5.2.3 Minimum Error Norm for Step Inputs to Stable Systems. . 94 5.2.4 Two-Degree-of-Freedom Controller . = ae OM SB MThe IMC Pilter yl 0h eke Mei ote aks to eraih a 98 5.3.1 Filter Form .. deceit 96 5.32 Qualitative Interpretation of the Filter Rimction 99 5.4 Robust Stability Ee eit oegldl 541 Norm-Bounded Uncertainty Regions |... 2.2... 101 5.4.2 General Uncertainty Regions jets 10? 5.5 Robust Performance .. . . sae 102 5.6 Summary of the IMC Design Procedure... 0... +++ ++ + 103 5.7 Applications wat +++ 104 5.7.1 Distillation Column Base Level Control 104 5.7.2. NMP Unstable Systems . . . . _ 107 5.8 References . ‘ 0 6 ISSUES IN SISO IMC DESIGN 113, 6.1 Implications of IMC for Classic Feedback Controllers... .. . . 113 6.1.1 General Relationships . hive sere seaxsase & 18 6.1.2 PID Settings for Simple Models . 114 6.1.3 PID Settings for a First-Order System with Dendtime |. . 121 614 Summary .. Lec ee iis 62 IMC Interpretation of Smith Predictor Controller... +. ss... 126 6.2.1 General Relationships = 126 63 6.2.2 Some Myths about the Tuning of Smith-Predictor Controllers128. 623 Robust Tuning of Smith-Predictor Controller for First- Order System with Dead 130 G24 Summary 6 Oo SEE. ha, Siege bse toiegew <2 5 1B Feedforward Control Pot Ste orc tee eps 6.3.1 Objectives and Structure au » 131 63.2 Design... . . + 132 63.3 Summary = 135 CONTENTS. 64 Cascade Control... . lat gyn Hel 180, GA Objectives, Structure, and Design |... 1... - 135 6.4.2 Implementation . + 138 643 Summary . ne - 138 6.5 References vars = 139) Part II: SAMPLED DATA SINGLE-INPUT SINGLE-OUTPUT SYSTEMS 141 FUNDAMENTALS OF SAMPLED-DATA SYSTEMS CON- TROL 143, 7.1 Sampled-Data Feedback Structure Poe ed eee ois. 7.2 IMC Structure heat ut 73 Formulation of Control Problem... .....+ss00s0000. 447 7.3.1 Process Model... « ds ti -. 49 732 Model Uncertainty Description 0... sss... es... M9 TA Internal Stabibity 5,-5.6. 6 +1. so Mealagee sig «bere «DEO 7.5 Nominal Performance . . . -- 161 7.5. Sensitivity and Complementary Sensitivity Function 181 7.5.2 Asymptotic Properties of Closed-Loop Response 153 7.5.3 Limitations on Achievable Performance Bois 165 7.54 Discrete Linear Quadratic (H}-) Optinat Control... .. . 158 7.5.5 Hao Performance Objective . . ie shy tg -+-5)9, 180) "76 Robust $tabilliby 9278 Sasen- eschti oS. afeoaye ee hee 05 e100 7.7 Robust Performance . . . eho fete, cee peat TTA Hy Performance Objective... 1.0... s cee eee es 16 7.2 Heo Performance Objective»... 6.0 seve eee + 162 78 Summary 162 79 References... 2.22... ok es e168 SISO IMC DESIGN FOR STABLE SAMPLED-DATA SYS- ‘TEMS 165 8.1 Nominal Performance ..... = 165 8.11 Hj-Optimal Control or eee ee) 8.1.2 Design of the IMC Controller q(2) ....-.. +--+ +» 168 8.2 The Discrete IMC Filter... . . = eat: 83 Robust Stability Sr eke sh PEE res ee 8.3.1 Filter Design ees iis 8.3.2 Effect of Sampling . . 176 84 Robust Performance 176 CONTENTS: 10 8.4.1 Filter Design... . guts 3! Lemay AT 842 Sampling Time Selection edge tee - 178 843 Example.......... - 178 8.5 Summary . 181 86 References . 182 SISO DESIGN FOR UNSTABLE SAMPLED-DATA SYSTEMS183 9.1 Parametrization of All Stabilizing Controllers 183 9.1.1 Internal Stability i + 183 9.1.2 Controller Parametrization 184 9.2 Nominal Performance 184 9.2.1 H}-Optimal Controller ora 185 9.2.2 Design of the IMC Controller (=). 188 9.2.3 Anti-aliasing Prefilter ....... 189 9.24 Design for Common Input Forms - 190 9.2.5 Integral Squared Exror (58) for Step Taps to Stable Sys- tems . : Jock. A192 9.3. The Discrete IMC Filter « 193 9.3.1 Filter Form 193 9.3.2 Qualitative Interpretation of the Filter Function = 195 9.4 Robust Stability - 198 9.5 Robust Performance . . Ba 198 9.6 Summary of the IMC Design Procedure...» - = 199 9.7. Application: Distillation Column Base Level Control... . .. . 200 9.8 References 2 a 201 Part II: CONTINUOUS MULTFINPUT MULTI-OUTPUT SYSTEMS 203 FUNDAMENTALS OF MIMO FEEDBACK CONTROL 205 10.1 Definitions and Basie Principles . = 205 10.1.1 Modeling gs = 205; TOO" Polea a, kat shart + cache = 206 10.1.3 Zeros 207 10.1.4 Vector and Matrix Norms . = = 208 10.1.5 Singular Values and the Singular Value Decomposition . . . 211 10.1.6 Norms on Function Spaces 215 10.2 Classic Feedback = 217 10.2.1 Definitions 217 10.22 Multivariable Nyquist Criterion - +217 103 104 10.5 10.6 aa 12 13 CONTENTS 10.2.3 Internal Stability . 10.24 Small Gain Theorem . . . Formulation of Control Problem Be 10.3.1 Process Model... . . « neha 10.3.2 Model Uncertainty Description 10.3.3 Input Specifications 5 a 10.3.4 Control Objectives... 6... + Nominal Performance ao 104.1 Sensitivity and Complementary Sensitivity Function . . 1042 Asymptotic Properties of Closed-Loop Response (System Type) - 10.43 Linear Quadratic (Hs-) Optimal Control 10.4.4 Hoo-Optimal Control Summary . . aye References 11 ROBUST STABILITY AND PERFORMANCE Robust Stability for Unstructured Uncertainty 11.1.1 Uncertainty Description... . 1112 General Robust Stability Theorem 11.1.3 Multiplicative Output Uncertainty .. . . 11.14 Multiplicative Input Uncertainty 11.5 Inverse Multiplicative Output Uncertaint 11.1.6 Example: Input Uncertainty for Distillation Column 11.1.7 Integral Control and Robust Stability . . Robust Stability for Structured Uncertainty 11.2.1 Uncertainty Description... . . 11.22 Structured Singular Value 11.23 Simultaneous Multiplicative Input and Output Uncertainty 11.24 Batch Reactor: Simultaneous Parametric and Unstruc- tured Uncertainty ‘ 11.25 Independent Uncertainty in the Transfer Matrix Elements 11.2.6 Condition Number and Relative Gain Array as Sensitivity Measures OEE el ie ataate Robust Performance é 11.3.1 Hoo-Performance Objective .......- 11.3.2 Multiplicative Output Uncertainty 11.33 Multiplicative Input Uncertainty . . . 11.34 Hy-Performance Objective . 11.35 Application: High-Purity Distillation « = 221 221 = 222 222 223 225 226 227 207 = 228 228 230 = 232 233 235, = 236 236 237 = 239 240 = 242 242 244 246 = 246 — CONTENTS 12 13, 114 Robustness Conditions in Terms of Specific ‘Transfer Matrices 11.4.1 How to find the LPT... .. 114.2 New Properties of ws. 5 « 114.3 Examples 11.5 Summary 116 References MIMO IMC DESIGN 12.1 IMC Structure 122 Conditions for Internal Stability 123 Parametrization of All Stabilizing Controllers 124 Asymptotic Properties of Closed-Loop Response 12.5 Outline of the IMC Design Procedure 12.6 Nominal Performance A 12.6.1 Assumptions i 12.62 Hy-Optimal Control for a Specific Input - 12.6.8 Hy-Optimal Control for a Set of Inputs 12.64 Algorithm for “Inner-Outer” Factorization 12.7 Robust Stability and Performance 12.7.1 Filter Structure ‘ 12.72 General Interconnection Structure with Filter « 12.7.3 Robust Control: Hoo Performance Objective 12.74 Robust Control: Hy-Type Performance Objective 12.8 Application: High-Purity Distillation . . : 12.9 Summary . 12.10Discussion and References . . PERFORMANCE LIMITATIONS FOR MIMO SYSTEMS. 13.1 Effect of Plant Gain 13.1.1 Constraints on Manipulated Variable 13.1.2 Disturbance Condition Number 13.1.3. Implications of cg for Closed-Loop Performance 13.14 Decomposition of d along Singular Vectors . 13.1.5 Summary 13.2 NMP Characterist 13.2.1 Zero Direction 13.2.2 Implications of Zero Direction for Achievable Performance 13.2.3 Summary 13.3 Sensitivity to Model Uncertaint 13.3.1 Sensitivity to Diagonal Input Uncertainty. = 301 - 301 275 278 281 285 287 290 293 + 293 = 295 + 296 297 298 302 306 309 310 311 313 315 ais 320 324 325 327 328 828 329 330 332 338 339 330 342 343 344 au 14 15 CONTENTS 13.3.2 Sensitivity with Different Controller Structures . . 13.3.3 “Worst-Case” Uncertainty Bhd : 13.3.4 Example... eae be er etean eae 13.3.5 Summary 134 References DECENTRALIZED CONTROL 14.1 Motivation 14.2 Definitions . 14.3 Necessary Conditions for Controllability. 14.3.1 Results 143.2 Proofs ae? 144 Stability Conditions - Interaction Measures... . 14.4.1 Necessary and Sufficient Stability Conditions 14.4.2 Sufficient Stability Conditions . a 14.43. Diagonal Dominance Interaction Measures . . 1444 Generalized Diagonal Dominance Interaction Measures . 14.4.5 The p-Interaction Measure ‘ 14.4.6 Interaction Measures for 2 x 2 Systems. 14.47 Examples i 14.5 Robust Performance Conditions . 14.5.1 Sufficient Conditions for Robust Performance 14.5.2 Design Procedure . . 14.5.3 Example . 146 Summary 147 References... . 346 348 M9 . 361 - 356 359 +6 359 - 361 » 362 - 362 365 = 367 = 368 . 370 372 373 374 = 315 - 376 - 378 379 382 383 388 389 Part IV: SAMPLED DATA MULTL-INPUT MULTI-OUTPUT, SYSTEMS MIMO Sampled-Data Systems 15.1 Fundamentals of MIMO Sampled-Data Systems 15.1.1 Sampled-Data Feedback a 15.1.2 Poles and Zeros 15.1.3 Internal Stability . . . 15.14 IMC Structure... . : 15.15 Model Uncertainty Description . 15.2 Nominal Internal Stability 15.2.1 IMC Structure 15.2.2 Feedback Structure 391 393 393, - 393 395 = 396 307 - 397 - 399 = 399 400 CONTENTS xiii 15.3 Nominal Performance... . 2006s eee0 esse eee ee 401 15.3.1 Sensitivity and Complementary Sensitivity Function... 401 15.3.2 Hj Performance Objectives 402 15.3.3 Hao Performance Objective 5 se 404 15.4 Robust Stability .......... woetiny 8 15.5 Robust Performance Potro Series 405 15.5.1 Sensitivity Function Approximation... . 405 15.5.2 Hg, Performance Objective ....... = 406 IMC Design: Step 1(Q) . 0... esc e eee dos kck ae 15.6.1 Hy-Optimal Control tat 8t- Sten nes 406 15.6.2 Setpoint Prediction SER es okdg tatu hiterae 418 15.6.3 Intersample Rippling... 0... 413 15.6.4 Inner-Outer Factorization . . sBit 5 ah Al 15.7 IMC Design: Step 2(F) o 0.0 cece cece eee eee ALO 15.7.1 Filter Structure . . . . wee 420 15.7.2. Robust Stability Interconnection Structure beh aal 15.7.3 Robust Performance Interconnection Structure... ... . 422 15.74 Robust Hz. Performance Objective... . . « 423 15.8 Mlustration of the Design Procedure... 6.200 ve eee ee ss A2d 15.8.1 System Description . 404 15.8.2 Design of Q . Lohenee Bawa - 426 15.8.3 Design of F . Srciamieeh 3 «ht ugg aS 15.9 Summary... 0... - 433 15.10References .... .. . 2434 Part V: CASE STUDY 435 16 LV-CONTROL OF A HIGH-PURITY DISTILLATION COL- UMN 437 6:1 Reatures.<.<'. ees 00) Gia, le em eaters Be O7 16.1.1 Uncertainty 437 16.1.2 Nonlinearity... . . Bipot ta aes sote 2 eeit ass 16.1.3 Logarithmic Compositions . ache es ASS 16.1.4 Choice of Nominal Operating Point . eee 9 16.2 The Distillation Column ...... Sackees trae) AG2A Moddalling (255). 2 e ete en Sp bh gee 2 440 16.2.2 Simulations... . . Braga 442 16.3 Formulation of the Control Problem... 0... 0... 0.0.5. 443 16.8.1 Performance and Uncertainty Specifications»... .... 443 16.3.2 Analysis of Controllers... . a der cr aa CONTENTS 16.3.3 Controllers = M44 164 Results for Operating Point Als... sss. s eee 445, 16.4.1 Discussion of Controllers . . Bah Kiet we M6 16.4.2 Conclusions : - 451 165 Effect of Nonlinearity (Results for Operating Point C) = 452 16.5.1 Modelling ....... ? oo ABR 16.5.2 p-Analysis tats San os AB 16.5.3 Logarithmic Versus Unscaled Compositions»... 1... 453 16.5.4 ‘Transition from Operating Point A to C . - 455; 16.6 Conclusions . . . . a Bhd 456 16.7 PisibrendeR «occ. $ Sai doe v weethas arena aagealt = 456 Appendix 459 References 469 Index ato PREFACE ‘The development of the state space approach in the early 1960s made it possi- ble for the first time to solve general linear multivariable control problems with relative ease. The new techniques seemed to hold high promise for practical ap- plications and attracted much interest — at least in the academic community. Fifteen years later several review papers (e.g., Foss, 1973; Kestenbaum et al., 1976) concluded that the impact on the industrial practice had been negligible. A number of possible reasons for this failure can be identified. For example, there was no smooth transition from the established and proven techniques and tools (PID controllers, Smith Predictor, Relative Gain Array, Decoupler, etc.) to the new ones. This led to great educational difficulties and from today’s perspec tive pointless debates in which the classic frequency domain approach was pitched against the modern state space techniques. More serious than this lack of under- standing which persisted for almost two decades was that the new techniques did not address some very fundamental issues which are at the very heart of feed- back design. For example, the concept of “nonminimam phase behavior” seemed forgotten. Also, the problem of model uncertainty and model error, which domi- nates process control, was only addressed via sensitivity analysis, whose validity is limited to infinitesimal model perturbations. After years of isolated but persistent criticism a new understanding and ap- proach started to emerge in the late 1970s. Some of the key theoretical con- tributions were the new formulation and parametrization of the optimal control problem by Youla and coworkers (1976), the definition of the Ha» control problem by Zames (1981), and the introduction of the structured singular value by Doyle (1982). These and other discoveries sparked again much enthusiasm and a flurry of research activity almost comparable to the 1960s. Only time will show the impact on the control practice. ‘The new approach does address many issues of practical importance like model uncertainty (robustness). Many research challenges remain however. New results appear monthly but the ideas have not matured enough to form a coherent pic- wi PREFACE ture. Moreover, the level of abstraction makes the new developments inaccessible to anybody but the researchers in this area. ‘This book attempts to bridge the gap between a good undergraduate train- ing in process control and the new arriving generation of robust control theory. ‘Though not a textbook, it is intended as a supplement for a graduate level con- trol course. Because the emphasis is on practical control system design methods, sections of the book should also be of interest to the industrial control engineer who wants to learn about the new powerful techniques. Some basic concepts from linear algebra and complex variables are assumed. ‘The book motivates the new theory with a series of typical process control examples: single-input single-output systems with time delay, multivariable dis- tillation column models, reactor models, etc. However, the applicability of the new techniques is not limited to process control. The emphasis is on the deriva- tion of new analysis and design tools. Preference is given to simple effective techniques over the most recent and most general theoretical discoveries. There is no doubt that the next few years will bring about a number of refinements which will affect some of the details in the book. We strongly believe, however, that the key ideas of robust control as presented here will profoundly impact pro- cess control understanding, teaching, and practice in the future. Indeed, some of the concepts have been applied already to industrial systems and some of the algorithms have become integral parts of industrial process control software. Many colleagues have influenced our thinking on the topic of robust pro- cess control. We wish to acknowledge in particular Reuel Shinnar, Coleman Brosilow, and John Doyle. Shinar made us first aware of the robustness prob- Jems. Brosilow convinced us of the advantages of the Internal Model Control structure, Doyle familiarized us with the new powerful theory which makes the practical robustness analysis of multivariable systems possible. ‘The initial outline for this book was developed with our late colleague Con- stantin Economou. We missed his high standards, his flair for communicating complex ideas, and his humor. The project was not the same without him. A number of present and former Caltech graduate students and research fel- lows have contributed either directly or indirectly to the material. In particular, we would like to mention Sigurd Skogestad, Dan Rivera, Dan Laughlin, Dan Lewin, and Clandio Scali. Brad Holt, Yaman Arkun, and Christos Georgakis have classroom-tested parts of this book and have provided us with valuable criticism. The contimions encouragement by our friends in industry was most important, We wish to acknowledge, in particular, Dave Prett and Carlos Garcia of Shell Development, Dave Smith and Bjorn Tyreus of DuPont, and John Hamer of Kodak. PREFACE xii Much of the research reported in this book was carried out at Caltech. ‘The extraordinary scholarly environment and supportive administration made this endeavor possible, The final sections of the book were completed while the first author enjoyed an appointment as the Gulf Visiting Professor of Chemical Engi- neering at Carnegie Mellon University. During the past year, the second author held a joint appointment with the Chemical and Nuclear Engineering Department and the Systems Research Center of the University of Maryland. The excellent, atmosphere at this institution allowed him to continue his work on the book. Over the years we received continuous financial support from the National Science Foundation, the Department of Energy, Shell Development, DuPont, and Kodak which enabled us to pursue our research objectives. Finally we would like to thank Glenn C. Smith for preparing most of the figures, Evangelos Petroutsos for assisting with TEX and IsIpX and Jan Owen who did such an excellent job typing the manuscript after suffering through three different word processing systems, M. Morari B. Zafirion NOMENCLATURE Abbreviations DIC Decentralized Integral Controllable Ic Integral Controllable IM Interaction Measure mmc Internal Model Control Is Integral Stabilizable ISB Integral Squared Error LFT Linear Fractional Transformation LHP Left-Half Plane MIMO Multi-Input Multi-Output MP Minimum Phase NMP Nonminimum Phase NP ‘Nominal Performance NS Nominal Stability PFE Partial Fraction Expansion RGA Relative Gain Array RHP Right-Half Plane RHS Right Hand Side RP Robust Performance RS Robust Stability sIso Single-Input Single-Output SSE Sum of Squared Errors ssv Structured Singular Value svD Singular Value Decomposition uc Unit Circle Symbols Biz) by(s)(O5(2)) ba(s)(b5(z)) € cum (8), e(=)(C(s),C(z)) eg(s) d(s) a(s) @(2)(d5(z)) e(s)(e"(z)) e(s) (3), f(2)(F(s), Fle) Als) G(s) GE(s) GHG) ha(s)(Ho(s)) H,(H3) H3(H;") np Im NOMENCLATURE Factor to preserve system type in q(2) and Q(z) Alllpass with the open RHP (outside the UC) poles of #(s)(0"(z)) Allpass with the open RHP (outside the UC) poles of u(s)(v"(2)) Field of complex numbers Field of complex matrices of dimension n x m SISO (MIMO) continuous and discrete classic feedback controllers Classic feedforward controller Effect of d'(s) on y(s) Disturbance entering py(s) transform of d(s) before (after) passing through 7(s) ‘Contimuous (discrete) error signal: Difference Ddetween y(s)(y*(2)) and r(3)(r*(2)) ‘Weighted error signal W2(s)e(s) SISO (MIMO) continuous and discrete IMC filters ‘Type 1 one-parameter discrete SISO IMC filter ‘Transfer function matrix used as argument of jin the RP SSV criterion for the continuous case The equivalent of G¥(s) for the discrete case; it depends on whether v =r or v=d ‘Transfer function matrix used as argument of un the RS SSV condition for the continuous (discrete) case SISO (MIMO) zero-order hold Hp") for n= m= 1 HEH") for m Subspace of LJ", containing the functions with analytic continuations in the RHP (includes all rational, proper, stable transfer function matrices) Subspace of L3"*" defined as the orthogonal complement of the L}”*" subspace which contains functions with analytic continuations inside the UC (H3"*" contains all rational, strictly proper, stable z-transfer function matrices). Imaginary part NOMENCLATURE 4.(8)(La(s)) ba(e)(Elw)) Le(s)(Ex(w)) Li(s(E(w)) Salo) En(w)(E, (w)) Lo(s)(o(w)) LA(L5) T(Z3") apy") P(s)(P(s)) B(0)(P(s)) P'(2) PH 2)(P*(2), Py2)) BPP), PH) Ba(s)Ba(2)(Pals), Palz)) Pa(s),Pa(s) Pm(8)sBm(8) du(s) Piel) Pur(s), Pie2)) a(s),9(z(Q(s), Q()) As), H2)(Q(s), O@)) Ges), 24(8)(G(8), ae(8)) as) Additive SISO (MIMO) uncertainty Bound on the additive uncertainty for the continuous (discrete) case (Bound on the) MIMO inverse multiplicative output uncertainty (Bound on the) MIMO input multiplicative ‘uncertainty Multiplicative SISO uncertainty Bound on the multiplicative SISO uncertainty for the continuous (discrete) case (Bound on the) MIMO output multiplicative uncertainty Lym (Lg) for m = Lp (ig) for m = 1 Space of nx m matrix valued functions that are square integrable on the imaginary axis (UC) ‘SISO (MIMO) plant; also denotes the process model in sections where p = p(P = P) is assumed SISO (MIMO) process model Zero-order hold discrete equivalent of p(s)(P(s)) without and with 7(s)(U(s)) included, respectively Zero-order hold discrete equivalents of f(s)(P(s)) without and with 7(s)(I(s)) included, respectively ‘Allpass with the NMP elements of f(s) and B'(z)(P(s) and P*(z)) respectively ‘True and model of the disturbance effect plant transfer function ‘True and model of the measurement device transfer function MP factors of f(s) and p*(z) (P(s) and P*(2)) respectively SISO (MIMO) continuous and discrete IMC controllers SISO (MIMO) continuous and discrete Ist step IMC controllers Disturbance rejection (setpoint tracking) Ist step and overall IMC controllers of the two-degree-of-freedom structure Feedforward IMC controller 1 Gn(2)(@u(2)) 7G) ge Re r(s)(r*(z)) r u(s)(u(z)) u(s)(v'(s)) u'(z) V(s)(V(2)) wr) va(s),04(2)(Va(s), Va(z)) um (s), vi¢(Var(s), Vez) w(s) wa(s)(w-(s)) Wi(s)(Wals)) u(s)(a(s)) v2) HEMT E) H(2) ‘Um(s) Greek Characters NOMENCLATURE SISO (MIMO) Hj-optimal IMC controller Factor included in g(z) and Q(z) to avoid intersample rippling Field of real numbers Residue Real part Contimuons (discrete) setpoint Sampling time Continuous (discrete) input to the plant External input r(s) or d(s) (normalized to impulse) transform of v(s) Square matrix whose columns consist of external inputs v(s)(v"(2)) Set of inputs u(s)(v/(s)) or v*(z) Allpass with the NMP elements of v(3),0"(2)(V(a), V(2)) respectively MP factors of u(s),0*(z) (V(s),V(2)) respectively Performance weight for the SISO case Performance weight for disturbance rejection (setpoint tracking) in the two-degree-of-freedom structure Input (Output) weights for the MIMO case Continuous plant (model) output s-transform of y(s)(J(s)) before and after passing through 7(s) respectively Ontput of measurement device a s)(C(s)) ‘SISO (MIMO) anti-aliasing prefilter A(a)(AG)) Block diagonal matrix containing the uncertainty for the continuous (discrete) case A Set of possible A's Au(s)(Aa(z)) Uncertainty block in A(s)(A(2)) (8) d RP block in A(s) e(s),£(s)(€(s), £(s)) (Nominal) sensitivity function for the continuous SISO and MIMO cases, respectively as) first step IMC nominal sensitivity function NOMENCLATURE @(2)(E"(2)) SISO (MIMO) discrete sensitivity function &-(8), B,(s)(éa(s), Bu(s)) Setpoint tracking (disturbance rejection) SISO and MIMO (approximate) continuous nominal sensitivity function for sampled-data systems E,(s)(E.(s)) (Nominal) MIMO continuous sensitivity function for sampled-data systems; it depends on whether v= d (approximation) or » =r (sy(6(@) Zero polynomial for the continuous (discrete) case n(s), H(s)(A(8), H(s)) (Nominal) Complementary sensitivity function for the continuous SISO and MIMO cases respectively as) first-step IMC complementary sensitivity function nt(2)(E(2)) SISO (MIMO) discrete complementary sensitivity function fies) (8) Setpoint tracking SISO (MIMO) continuous nominal ‘complementary sensitivity fanction for sampled-data systems Co) (Minimized) condition number a Disturbance condition number kp Rijnsdorp IM ACA) RGA matrix of the matrix A (A) i eigenvalue of A dA) éj-element of A(A) H(A) oF wa) SSV of A for same specific set A 0 Set of possible plants #(s)((2)) Pole polynomial for the continuous (discrete) case x(w)("(w)) Uncertainty region on the Nyquist plane for the continuous (discrete) case oA) Spectral radius of A a(A) ‘Maximum singular value of A ata) Minimum singular value of A ai(A) i singular value of A 9) Factor included in f(z) to satisfy the requirements on the IMC filter oe Closed-loop system bandwidth % Sampling frequency Superscripts H Complex conjugate transpose of a matrix in ‘Transpose of a matrix 1 Orthogonal complement of a function space Special Notation F(A,A) cy 2} Well Welle ele + Meo, We he {o(s)}+({9(2)}+) {olo)}-ate)}-) > NOMENCLATURE Matrix resulting from the application of a linear fractional transformation on A, A with A in the feedback path Laplace transform of (-) Z-teansform of (-) Norm on ¢™™ Frobenius norm on ¢™™™ LR or 13" norm ‘o-norm of a function Operator norm induced by | lp Projection of an L3(L3") function 9(s)(9(2)) on HE(H3") Projection of an £3(L3") function 9(s)(o(z)) on (E13)*((H5") For all Such that ‘We also wish to remind the reader that the following statements are equivalent: ASB AitB B is sufficient for A Boa Boonly if A A is necessary for B not A => not B Chapter 1 INTRODUCTION ‘The main ideas covered in this book are motivated and put in historical perspec- tive. A new controller parametrization is outlined and the concept of “robustness” is explained. The scope of the book is defined and hints for its study are provided. 1.1. The Evolution of Control Theory Feedback control mechanisms have been used for millenia. The first applications of the feedback principle (Mayr, 1970) can be traced back to ancient Greece: the water clock of Ktesibios (ca. 300 B.C.) employed a float regulator. The first automatic feedback controller used in an industrial process was James Watt's fiyball governor invented in 1769 for controlling the speed of a steam engine. As these centrifugal governors were refined, they developed some serious problems, then referred to as “hunting” and now known as “instability.” This in turn prompted the first mathematical analysis of a feedback system via differential equations by Maxwell (1868). During the same period Vyshnegradskii (1877) formulated a mathematical theory of regulators. The invention of the negative feedback amplifier at Bell Laboratories by Black prior to World War II marks another milestone in the use of feedback. ‘The observed stability problems (“singing”) were explained through the frequency domain analysis techniques by Bode and Nyquist (Nyquist, 1932; Bode, 1964; Black, 1977). ‘Throughout the 1940s and 1950s feedback design by trial and error was pre- dominant and even today this approach plays a major role. ‘Typically, perfor- mance specifications are defined initially and then the controller design is modified iteratively until the best compromise between the frequently conflicting objectives is reached. Every design step is followed by an analysis step and the iterations are guided by experience and rules of thumb. In the late 1950s Newton, Gould, and Kaiser (1957) showed that if the control objectives are formulated in terms of the integral square error (ISE) the “optimal” feedback controller can be found 1 2 CHAPTER 1. INTRODUCTION “analytically” — ice, directly and without trial and error. This approach, which is based on the Wiener-Hopf factorization technique, was subsequently greatly generalized, in particular by Kalman. The deficiency of these “synthesis” techniques is that typical practical per- formance objectives are usually much more complex than ISE, Frequently the designer wishes to impose constraints on the closed-loop system response chat- acteristics like overshoot, rise time, and decay ratio. Also a certain “robustness” against changes in the dynamic characteristics of the plant to be controlled desirable. Often it is possible to achieve these attributes indirectly by frequency weighting the ISE and by adding to the objective a penalty for excessive move- ments of the manipulated variable. For that purpose the weights and penalties have to be chosen again iteratively by trial and error. Thus, for a practical design the advantages of such an “optimal control synthesis” procedure over an ad hoc adjustment of controller parameters and a test via simulation can become quite small. ‘The ISE or, in the more general context, the Ho-optimal control formulation dominated the literature for about 20 years. In the late seventies three major and in some sense related discoveries started a new era of feedback control theory. ‘* Youla and coworkers (1976) showed that it is possible to parametrize all sta- bilizing controllers for a particular system in a very effective manner: when searching for a controller with specific properties one can simply and without Joss search over the space of all stable transfer functions. The parametriza- tion guarantees that the resulting feedback controller automatically yields a closed-loop stable system. Thus, the search for a good controller is greatly simplified. ‘¢ Zames (1981) postulated that measuring performance in terms of the oo- norm rather than the traditional 2-norm (ISE) might be much closer to the practical needs. This ushered in the era of Ha. optimal control. # Doyle argued that model uncertainty is often described very effectively in terms of norm-bounded perturbations. For these perturbations and the Ha. performance objective he developed a powerful tool (the structured singular value) for testing “robust stability” (j.e., stability in the presence of model ‘uncertainty) and “robust performance” (ie., performance in the presence of model uncertainty). This is probably the primary motivation for the modern co-norm objective. In parallel to these developments in mathematics and electrical engineering a new type of algorithm exemplified by Model Algorithmic Control (Richalet et al., 1.2, CONTROLLER PARAMETRIZATION: THE IMC STRUCTURE, 3 1978) and Dynamic Matrix Control (Cutler and Ramaker, 1979) was invented in the process industries and applied successfully to complex process control prob- ems. Though these algorithms had a heuristic basis Garcia and Morari (1982) discovered that some of the modern robust control characteristics had been in- corporated in them in an ad hoc fashion. Thus, the time was ripe to put these empirical algorithms on a firmer footing and to have them benefit from the rich new theory. The objective of this book is to do that while retaining the features which make these algorithms easy to understand and easy to adjust on-line. At times this required mathematical generality to be sacrificed. 1.2 Controller Parametrization: The IMC Structure Consider the block diagram shown in Fig. 1.2-1 where the control system is shaded: it imcludes the two blocks labeled controller and model. The control system has as its inputs the setpoint and process output (measurement) and as its output the manipulated variable (process input). Let us discuss qualitatively the advantages of such a structure over the classic feedback structure shown in Fig. 1.2-2. ‘The effect of the parallel path with the model is to subtract the effect of the manipulated variables from the process output. If we assume for the moment, ‘that the modelis a perfect representation of the process, then the feedback signal is equal to the influence of disturbances and is not affected by the action of the manipulated variables. Thus, the system is effectively open-loop and the usual stability problems associated with feedback have disappeared. The overall system is stable simply if and only if both the process and the IMC controller are stable. Moreover, the IMC controller plays the role of a feedforward controller and can be designed as such. But the IMC controller does not suffer from the disad- vantages of feedforward controllers: it can cancel the influence of (unmeasured) disturbances because the feedback signal is equal to this influence and modifies the controller setpoint accordingly. If the model does not mimic the dynamic behavior of the process perfectly then the feedback signal expresses both the influence of (unmeasured) disturbances and the effect of this model error. The model error gives rise to feedback in the true sense and leads to possible stability problems. ‘This forces the designer to “detune” the ideal feedforward controller for “zobustness.” ‘The controller form described here is a special case of the Yorla- parametrization mentioned above. It is inherent in all “model predictive” control schemes, in particular Model Algorithmic Control and Dynamic Matrix Control. 4 CHAPTER 1. INTRODUCTION 1.3 Robustness Regardless of what design technique is used, controllers are always designed based ‘on (necessarily incomplete) information about the dynamic behavior of the pro- cess. This information (i.e., the “model”) can have the form of a system of coupled partial differential equations or be simply the process gain and the settling time experienced by the plant operator. The accuracy of this information varies but is never perfect. Moreover, the behavior of the plant itself changes with time (feedstock changes, catalyst activity changes, etc.) and these changes are rarely captured in the models. It is most desirable that the controller be insensitive to this kind of model uncertainty, ie., the controller should be robust. ‘Though the design objective robustness seems most practical and reasonable it is essentially absent from the control literature from about 1960 to 1980. This is probably one of the reasons why the design techniques developed in this time period have had negligible effect on the industrial control practice. Since the late 1970s robustness has become a major objective of control research. This book will, demonstrate the importance of robustness considerations for process control and will propose design techniques which include robustness as one of the objectives. 1.4. Scope of Book ‘The book has several goals: ‘to alert the reader to the key role model uncertainty and robustness play in the design and successful operation of feedback control systems. ‘to outline the basic ideas behind the most recent theoretical advances for addressing robustness systematically in feedback design. ¢ to describe simple and effective analysis and design techniques for robust feedback controllers. ‘The different parts of the book aim at different groups of readers: « the first-year graduate student who wants to broaden his/her horizon, gain some basic insights into feedback control, and get a taste of the direction of the new theory. « the beginning researcher in the area of robust control who wishes to master the transition from the undergraduate process control course to the mono- graphs by Vidyasagar (1985) and Francis (1987) as well as the current re- search papers in this area. .5. SOME HINTS FOR THE READER 5 # the industrial practitioner who needs easily applicable and proven controller tuning techniques and who is interested in learning about the basic dos and don’ts of multivariable control. Depending on the reader's objectives and background some chapters might be af mote interest than others. Chapters 2 through 4 discuss the basic trade-offs in single-input single-output (SISO) feedback control systems and how they can be addressed in the Internal Model Control framework. These chapters are essential ‘or understanding the rest of the book. Chapter 5 generalizes the results to anstable systems and can be skipped by the application-oriented reader. Chapter 3 proposes some very effective PID tuning rules and discusses more general SISO sontrol structures like Smith predictor, cascade and feedforward control systems. Chapters 7 through 9 deal with sampled-data systems and mimic Chapters 1 ‘through 6, In the derivations particular emphasis is placed on the behavior of the continuous output — e.g., provisions against intersample rippling are built into the controllers. Chapters 10 and 11 cover basic aspects of robustness of multi-input: multi- output (MIMO) systems. These concepts are the key to understanding the un- sual phenomena observed in the control of MIMO systems. The MIMO design technique proposed in Chapter 12 is complex and tentative. Undoubtedly, more affective techniques will become available in the future. Chapter 13 discusses the system properties which limit the achievable performance of MIMO systems. A knowledge of these properties helps when screening designs according to their op- erability characteristics. Chapter 14 deals with the design of controllers for MIMO systems with a restricted information structure — e.g., multiloop controllers. Chapter 15 extends the discussion of Chapters 10 through 12 to sampled-data systems. The book concludes with the detailed report on an application to a high-purity distillation column (Chapter 16). 1.5 Some Hints for the Reader The book assumes that the reader has mastered the material covered in a typical undergraduate control text like the one authored by Stephanopoulos (1984) or Franklin et al. (1986). Furthermore some knowledge of linear algebra (e.g, Strang, 1980) and complex variables (eg, Churchill & Brown, 1984) is required. Finally, some familiarity with linear operator theory (e.g., Desoer & Vidyasagar, 1975; chapters 1 through 6 of Ramkrishna & Amundson, 1985) is advantageous, but not necessary. 6 CHAPTER 1. INTRODUCTION SetPoint Feedback Sigaal Fig, 1.2-1, Internal Model Control Structure, Chapters 5, 9 and 13 and in particular the proofs in these chapters are mathe- matically more demanding and can be skipped at first reading. The chapters on sampled-data systems (7 through 9 and 16) assume that the corresponding mate- rial on continuous systems has been studied first. Also the study of the chapters ‘on MIMO systems requires the prior understanding of the SISO concepts. In principle, the SISO material follows as a special case from the MIMO material, For tutorial reasons it is treated separately. With the exception of this first chapter all references are collected at the chap- ter end for continnity. Most chapters contain one or several summary sections, usually located immediately before the reference section. This summary section allows the reader to review the main concepts. Often, mastering this summary section is sufficient for the application of the basic techniques covered in the chap- ter, The equation numbers in the summary section refer to the equation numbers in the main body of the text. ‘The end of proofs and examples is marked with the symbol Q. 15. SOME HINTS FOR THE READER Influence of disturbances SetPoint Fig, 1.0.2. Classic Feedback Structure. Part I CONTINUOUS SINGLE-INPUT SINGLE-OUTPUT SYSTEMS Chapter 2 FUNDAMENTALS OF SISO FEEDBACK CONTROL After a review of some basic definitions, the controller design problem will be for~ mulated in general terms. All controller design procedures are based on models of one form or another. These models are necessarily inaccurate. It is of great practical importance that the controller performs well even when the dynamic behavior of the real process differs from that described by its model. In order to accomplish this objective, not only must a process model and the performance specifications be provided for the design, but also some indication of the model accuracy should be made. The attribute “robust” will be used for a property ‘which holds not only for the model but also in the presence of model uncer- tainty. Mathematical conditions for robust stability and robust performance will be derived. 2.1 Definitions The block diagram of a typical classic feedback loop is shown in Fig. 2.1 Here ¢ denotes the controller and p the plant transfer function. ‘The transfer function py describes the effect of the disturbance d’ on the process output y. ‘The measurement device transfer function is symbolized by Py. Measurement noise n corrupts the measured variable yy. ‘The controller determines the process input (manipulated variable) u on the basis of the error e. The objective of the feedback loop is to keep the output y close to the reference (setpoint) r. Commonly we will use the simplified block diagram in Fig, 2.1-1B. Here d denotes the effect of the disturbance on the output. Exact knowledge of the output y is assumed (Pq = 1,n = 0) In general, the transfer functions will be allowed to be rational or irrational — ice., they may include time delays. In order to be physically realizable, the transfer functions have to be proper and causal nu 2 CHAPTER 2, FUNDAMENTALS OF SISO FEEDBACK CONTROL Figure 2.1-1. General (A) and simplified (B) block diagram of feedback control system. Definition 2.1-1. A system g(s) is proper if lims4..9(s) is finite. A proper system is strictly proper if lim,_.4. |g(s)| = 0 and semi-proper if lim, 4. |g(s)| > 0. All systeras which are not proper are improper. A system g(s) is improper if the order of the numerator polynomial exceeds the order of the denominator polynomial and proper otherwise. An improper system cannot be realized physically because it contains pure differentiators. Definition 2.1-2. A system requiring prediction (e*#) is noncausal A system which does not require prediction is causal. Some care has to be used in determining if a system is causal or not. A time delay in the denominator of a transfer function does not necessasily imply lack of causality. For example, ee BU sree a hy 8) = GQ) = Fea Tew is causal becanse the present output y is determined solely by past values of y and the present value of u: Hee) = lea) + u(a)) —TGenerally an improper qyalem i also referred to as noncausal. In thie book we will reserve soncatal to denote prediction 22, FORMULATION OF CONTROL PROBLEM 13 -100 360 aos ae aot 20) Figure 2.1-8. Phase behavior of two nonminimum phase (NMP) allpass elements Definition 2.1-3. A system 9(s) is nonminimum phase (NMP) if its transfer function contains zeros in the Right Half Plane (RHP) or time delays or both. Otherwise a system is minimum phase (MP). Systems with an odd number of RHP zeros display inverse response charac- teristics: the initial direction of the step response is opposite to the direction of the final steady state. ‘The origin of the term “minimum phase” can be best understood in the context of stable systems: if'a system g(s) is MP, then there is no other system g/(s) which has the same magnitude (Ig(‘22)| = lg(is)| Vu) but a smaller phase lag. In other words, a NMP system 9(3) does not have the smallest phase lag that is possible for a system with this magnitude (g(iw)|). For example, the alipass systems +2 st? a(s) = gals) have lou ie)| = lao(iw)| = 1 ‘Their phase behavior is shown in Fig. 2.1-2. Clearly there is another system with the same magnitude (for example, g'(s) = 1) which has a smaller (zero) phase Jag. Therefore, g(s) and go(s) are NMP. 2.2 Formulation of Control Problem For any design procedure to yield a control algorithm which works satisfactorily in a real environment, the following must be specific ¢ process model 4 CHAPTER 2. FUNDAMENTALS OF SISO FEEDBACK CONTROL ‘© model uncertainty bounds «* type of inputs (.e., setpoints and disturbances) +» performance objectives Omission of any of these four items invariably leads to controllers which fail in practice. Bvery design or tuning procedure is centered around a process model whose complexity can vary. The experience of the operator who “knows” how the plant responds to certain inputs is the simplest kind of model. At the other end of the scale would be, for example, a coupled system of nonlinear partial differential equations. Neglecting model wncertainty leads to controllers which are too “tight” and which are likely to become unstable in the real operating environment. It is physically impossible to design controllers which work well for all types of setpoint changes and disturbances. The designer has to decide which inputs are most important, and most frequent, and accept inferior performance if the inputs encountered during the operation are not exactly equal to the inputs assumed for the design. Finally, the designer has to specify what is meant by “good performance” for the particular problem at hand. What is good in one case might be entirely unacceptable in another. 2.2.1 Process Model Chemical processing systems are often most conveniently described by infinite dimensional linear time invariant models (transfer functions with delays). Thus, aay book on process control must discuss how to design controllers based on this type of model. However, a rigorous theoretical treatment of time-delay systems is very complex and more of mathematical than of practical interest. The reason is that there are extremely few true time-delay processes encountered in practice. In general, a process modeled with a delay is actually a finite dimensional system of very high order whose dynamic behavior can be approximated by a delay. The delay in this case is simply a convenient modeling tool which does not have to be treated rigorously in the design procedure. Also, as we will argue in more dotail in the next section, all models are inadequate at high frequencies. In the low-frequency range, the delay can be approximated arbitrarily well by a finite dimensional (j.e., delay-free) model. In the high-frequency range the difference between the delay and the finite dimensional approximation can be treated as “model uncertainty” and the controller can be designed accordingly. (One key objective of this book is to show how to account for model uncertainty in the controller design procedure.) Strictly speaking, most results obtained in this book are only applicable to systems described by rational transfer function models. In a less rigorous manner 22. FORMULATION OF CONTROL PROBLEM 15 ‘we will extrapolate the results to systems involving time delays, in particular when this allows us to obtain compact analytical expressions. The mathematically precise reader should interpret the time delays in our equations as a shorthand notation for a very high-order Padé approximation. 2.2.2 Model Uncertainty Description Linear time invariant models of the type used throughout this book describe actual plant dynamics only approximately. ‘The “model uncertainty” can have several different sources. Most important, real processes are nonlinear. If the process model is obtained via linearization, then it is accurate only in the neigh- borhood of the reference state chosen for the linearization. In other cases the process might be represented quite accurately by a linear model. However, different operating conditions could lead to changes in the pa- rameters in the linear model. For example, increased throughput and flowrates usually result in smaller deadtimes and time constants. In the two cases above, the sources and the structure of the “uncertainty” are known quite accurately. However, there is always some “true” uncertainty even when the underlying process is essentially linear: the physical parameters are never known exactly and fast: dynamic phenomena (e.g., valve dynamics) are usually neglected in the model, Therefore, at high frequencies, even the model order is unknown Uncertainty can be described in many different ways: bounds on the parame- ters of a linear model, bounds on nonlinearities, frequency domain bounds, etc. ‘To account for model uncertainty we will assume in this book that the dynamic behavior of a plant is described not by a single linear time invariant model but by ‘a family TI of linear time invariant models. This is a somewhat primitive uncer- tainty description especially when the effect of nonlinearities is to be captured, but it is the only feasible approach at present. ‘Throughout this book the family I of plants will be defined in the frequency domain. We will assume that the transfer function magnitude and phase at a particular frequency w is not confined to a point but cam lie in a region 7(w) on the Nyquist plane. In general, the region can have a very complex shape as the following examples illustrate. Assume, for example, that frequency response experiments allowed the de- signer to establish an upper and a lower bound on both the magnitude [p(iw)] and the phase arg {p(iw)} of the real plant. At a particular frequency these bounds give rise to the sector shaped region +(w) in Fig. 2.2-1. 16 CHAPTER 2. FUNDAMENTALS OF SISO FEEDBACK CONTROL ( © Figure 2.9-1. Uncertainty regions resulting from gain and phase bounds (A) and parametric bounds ©). a aper D c Figure 2.2-1, Uncertainty “band” (C). Uncertainty disk for aditive uncertainty (D). On the other hand, step response experiments might have yielded a first-order model with deadtime (f(s) = ke~*"(rs-+1)~) with upper and lower bounds on the parameters k, @ and r. The resulting uncertainty region is shown for a particular frequency in Fig. 2.2-1B. The union of all regions x(w) constitutes the family of possible plants II and can be viewed as a “fuzzy” Nyquist plot or Nyquist band (Fig. 22-10). Regions of the general shape seen in these examples have complex mathe- matical descriptions and are very difficult to deal with in the context of control system design. Therefore, in most of this book we will assume the regions to be disk shaped with radius Z(w) (Fig. 2.2-1D). Any complex region can be ap- proximated by a disk with more or less conservatism (see Fig. 2.2-1A and B). Algebraically, the family TT of plants described by the disk is defined by: TL = {p: [p(iw) — Bliw)| < Fa(w)} (22-1) Here p(iw) is the nominal plant or the model defining the center of all the disk- shaped regions. Any member of the family II satisfies (iw) = Bi) + bale) (22-2) with 5 LaCie] < Falwe) (22-3) 2.2. FORMULATION OF CONTROL PROBLEM 7 Equation (2.2-2) is referred to as an additive uncertainty description and (2.2-3) states a bound on the allowed additive uncertainty. If we define Em (it (2.2-4) and (2.2-8) then the family II (2.2-1) can be represented as ofp, Win) Ba S nig fp: Teh sh (2.2 -6) ‘Thus, any member of the family II satisfies (iat) = Pliw)(1 + Elita) (22-7) with Venta) < baw) 02-8) Equation (2.2-7) is referred to as a multiplicative uncertainty description and (2.2-8) states a bound on the allowed multiplicative uncertainty. A typical plot of Z,(w) is shown in Fig. 2.2-2. The uncertainty usually increases with frequency and eventually exceeds or becomes equal to unity. The reason is that our models tend to describe well the steady state and low-frequency behavior of processes but become inaccurate for high-frequency inputs. If a zero is located on the imaginary axis at w* for some p € II then p(iu*) = 0 and [fm(iu*)| = 1. If &q(w) > 1 for w > w*, then the set TI allows zeros of p(s) sufficiently far from the the origin “to cross” the imaginary axis. Also, for w > w* the disk-shaped regions include the origin which implies that the phase is completely unknown, On the other hand, if a pole is located on the imaginary axis at w* for some PET then |p(i")| = 00 and therefore &(u*) = En(wo") = oo. Thus the additive and multiplicative uncertainty are unsuitable to describe sets of plants where poles can cross the imaginary axis or equivalently, the number of RHP poles can vary. Figure 2.2-3 shows a block diagram representation of the two types of uncer- tainties, 18 CHAPTER 2, FUNDAMENTALS OF SISO FEEDBACK CONTROL tee al Figure £.2-8. Typical form of multiplicative uncertainty bound la. Figure 2.2-9. Additive (,) and multiplicative (fq) uncertainty, 22. FORMULATION OF CONTROL PROBLEM 19 igure 2.9-4, Weight w transforming normalized input oto physical input ». 2.2.3 Input Specification We refer to all external signals entering the feedback loop at some point as inputs. ‘The symbol v will be used to symbolize these inputs which could be setpoint changes or disturbances. For the control system design the inputs have to be specified. The inputs will be normalized for notational convenience (Fig. 2.2- 4), It will be assumed that an input v entering the control loop is generated by ‘passing the normalized input v' through a transfer fnetion block w(s), sometimes referred to as an input weight. Our design procedures will be able to deal with two types of normalized input specifications: Specific input (impulse): v(s)=1 2-9) ‘Set of bounded inputs (all inputs with 2-norm bounded by unity): ot wR S [vat <1} (22-10) Very often in process control we know quite well what type of inputs to expect. For example, setpoint changes usually have the shape of steps or ramps, distur- ances can often be modeled as steps, entering a lag. If the control system design is to be carried out to accommodate a specific expected input v, such as a step setpoint change, then the designer specifies the input weight w to be an integrator (a/s). Specific input: vew! w (22-1) However, sometimes the closed-loop performance can be quite sensitive to the choice of input; if the input assumed for the design is not exactly equal to the input encountered in practice the performance can deteriorate significantly. Then it is more meaningful to consider a set of inputs containing the input: which is encountered most frequently together with other “similar” inputs. According to (2.2-10) we will consider the 2» CHAPTER 2, FUNDAMENTALS OF SISO FEEDBACK CONTROL wt 5 2 Seas le ele el Figure 2.9-5. Typical form of input weight w. Set of bounded inputs: na (8p. 1 jolie) e WIR | EL = a5 Lalouanl & $2 (22-12) Note that the frequency integral used in (2.2-12) for the 2-norm is equal to the time integral in (2.2-10) through Parseval's theorem. ‘The expression (2.2-12) can be interpreted as follows: if the spectrum of the input v is narrow and concentrated near «* (i.e., the input looks almost like sin w*t) then the input power (amplitude) is limited to jo(iw*)|. A typical input weight is shown in Fig. 2.2-5 indicating that high-frequency inputs are expected to have small amplitudes. Assume, for example, that we expect unit step inputs v. Then we can define a weight w such that a step lies in the set (2.2-12). For example the weight ot uw) = ag 9>0 (2.213) has the desired characteristics: the input ' _ (22-14) satisfies ‘A 26 Ww = 5 [apa (22-15) and gives rise to a step u when passing through the weight (2.2-13). However, the characterization (2.2-12) with the weight (2.2-13) contains also many other 22, FORMULATION OF CONTROL PROBLEM a signals v which are “similar” to steps. For example, consider the input We v a>0a¢B (2.2-16) Sta which also satisfies (2.2-10). After passing through the weight the signal becomes @s+8 B(s+a)s which is a step modified by a lead (for a > 8) or a lag (for @ < A). Thus, if the controller is designed for the set (2.2-12) with weight (2.2-13) it will work well not only for steps but also for “modified steps,” for example of the form (2.2-17). In order for a physically realizable controller to exist, the external signals which are assumed to enter the control loop have to be bounded. Unbounded external signals can only be corrected by unbounded control actions. Therefore, they do not give rise to a meaningful control problem formulation. In general, the effect of the disturbances on the output (¢) is bounded and the disturbance, transfer function pq (Fig. 2.1-1A) is stable. If the plant is open loop unstable and a (bounded) disturbance enters at the plant input, its effect on the output is usually unbounded and pg is unstable. In this special case the unstable poles of a are a subset of the unstable poles of the plant p. Therefore, a controller with bounded control action can be designed, which provides assymptotic disturbance rejection. (22-17) 2.2.4 Control Objectives The ultimate objective of control system design is clearly that the controller works “well” when implemented on the real plant. This goal can be best understood if ‘we decompose it into a series of subobjectives. Because we assume that the model is at least an approximate description of the true plant it is reasonable to require stability when the controller is applied to the plant model Thus, the minimal requirement on the closed-loop system is nominal stability. In Sec. 2.3 we will introduce the concept of internal stability to derive the conditions for nominal stability which must be satisfied through appropriate controller design. As a measure of performance we will generally use the Integral Square Error (ISE) because of its mathematical convenience. In Sec. 2.4 we will formulate the optimal control problem for minimizing the ISE. It is essential that our knowledge about the model uncertainty be incorporated into the controller design procedure. Otherwise the controllers are bound to fail in the real world where the actual plant behavior can be quite different from that 2 CHAPTER 2, FUNDAMENTALS OF SISO FEEDBACK CONTROL Figure 23-1. Block diagram for discussion of internal stability of the model. We will require that the controller be designed such that the closed- loop system is stable and meets the performance specifications for all members of the family II of possible plants defined in Sec. 2.2.2. If a closed-loop property holds for a family of plants we will refer to it as robust. ‘Thus, our controller design objectives are robust stability and robust performance. In Secs. 2.5 and 2.6 we will derive conditions for robust stability and robust performance, respectively. 2.3. Internal Stal ity ‘The signals between the blocks constituting a control system are subject to (pos- sibly very small) errors. In practice it cannot be tolerated that these small errors lead to unbounded signals at some other location in the control system. This motivates the following definition: Definition 2.3-1. A control system is internally stable if bounded signals injected at any point of the control system generate bounded responses at any other point. Definition 2.3-2. A linear time invariant control system is internally stable if the transfer functions between any two points of the control system are stable — ie, have all poles in the open Left Half Plane (LHP). In a control system many different points can be selected for signal injection and observation but most of the choices are equivalent for checking internal sta- bility, For example, for the system in Fig. 2.3-1, y and e differ only by a bounded signal (r) and their observation reveals the same information about internal sta- lity. Also, from the point of view of internal stability, the effect of d and r on wis equivalent. Simple arguments of this type reveal that there are only two “independent” outputs, which can be chosen as y and u, and two “independent” inputs, which can be chosen as rand w/. Thus, the classic feedback system is 2.4, NOMINAL PERFORMANCE 3 stable if and only if all elements in the 2 x 2 transfer matrix in (2.3-1) have all their poles in the open LHP. ‘()-( ue) (3) (23-1) uw) Nii Tie] ht! It should be noted that the concept of internal stability is more complete than the usual stability concept of undergraduate textbooks where system stability is checked by examining the roots of the characteristic equation 1+pc = 0. Consider the example mst1 ae (23-2) 1 a (23-3) for which the characteristic equation 1 L+pe= 1+ =0 (23-4) has a single root at s = -2 which would indicate stability. On the other hand, the transfer matrix in (2.3-1) evaluated for this example ()-( B)@) es-» shows that any bounded input wu leads to an unbounded output y. Thus, the system is not internally stable. Note that the RHP zero of the controller (2.3-2) cancels exactly the RHP pole of the plant (2.3-3) when (2.3-4) is formed. The internal stability concept clarifies the fact that even ezact cancellation is not enough to guarantee the stability of the system, If both p and c are stable then the unstable poles in (2.3-1) can arise from (1+ pe)~! only. In this special case it is necessary and sufficient for internal stability that all the roots of the characteristic equation 1 + pc = 0 are in the open LHP. 2.4 Nominal Performance The most basic objective of a feedback controller is to keep the error between the plant output y and the reference r small when the overall system is affected by external signals r and d. In order to quantify performance, a measure of % CHAPTER 2. FUNDAMENTALS OF SISO FEEDBACK CONTROL “smallness” for the error has to be defined. Furthermore, the set of external input signals has to be specified for which the error is to be made small. Control system design techniques differ in the way they measure “magnitude” and how they define the permissible set of external inputs. Two particularly popular approaches are outlined in Secs. 2.4.4 and 2.4.5. 2.4.1 Sensitivity and Complementary Sensitivity Function It we st 71= P= 1, then the son important relaionsipn betwen the fps and outputs in Fig. 2.1-1A are 1 Tre Re pols) 4 S ran “Trp eae where the error ¢ is defined as e = y—r ‘The sensitivity function e(s)? relates the external inputs d ~ r to the error It also expresses the effect of the disturbance d on the output y. The sensitivity ¢(s) is of primary importance in judging the performance of a feedback controller. It is desirable to make «(s) as “small” as possible. If pe is strictly proper (which is always the case for physical systems) then lim, pe=0 @4-3) ‘This implies that Ji tit = Jin 4-4 ‘Thus |e| can be made small only over a finite frequency range. Physical limitations do not allow “perfect control” (¢ = 0). A typical plot of [e| is shown in Fig. 24-1. The frequency wp at which |e| exceeds 1/3 will be called the system band- width, lew) <1/V2 Ww |w] w small (24-23) That is, the loop gain |pc| has to be shaped to fall above the performance weight fw. 2.5 Robust Stability We wish to derive conditions for the robust stability of the family IT of plants defined by (2.2-6). For this purpose we will employ the Nyquist stability criterion. 2 CHAPTER 2, FUNDAMENTALS OF SISO FEEDBACK CONTROL Figure 2.5-1. Diagram for graphical derivation of robust stability and robust performance condi- tions. Let us denote by N(k, 9(s)) the net number of clockwise encirclements of the point (k,0) by the image of the Nyquist D contour under o(s). We will assume that all peII have the same number (n) of RHP poles. Then we have robust stability for a specific controller ¢ if and only if (-1,pe) = vp ell (2.5=1) In particular it is necessary that the nominal plant be closed loop stable: N(-1,pc) = -n (25-2) Assuming that (2.5-2) holds, (2.5-1) will hold if and only if the Nyquist band which comprises all p € II does not include the point (-1,0). From the simple geometric argument in Fig. 2.5-1 it is clear that this will be the case if and only if the distance of je from the point (-1,0) — i.e., |1 + fc], exceeds the disk radius Be(ite Emu): [1+ pe(iw)| > [Beliv)[En(w) Yar or IT potty) = Viliw)[im(w) <1 Ya (2.5 —3) where f= fe(1 + fe)"!. We a proven the following theorem. 25. ROBUST STABILITY 33 ‘Theorem 2.5-1 (Robust Stability). Assume that all plants p in the family II (22-6) have the same number of RHP poles and that a particular controller ¢ stabilizes the nominal plant j. Then the system is robustly stable with the controller c if and only if the complementary sensitivity function ij(8) for the nominal plant p satisfies the following bound Unlloo sup Filinie)] <1 (25-4) Let us clarify what it means that Thm. 2.5-1 is not only sufficient for robust stability but also necessary. If condition (2.5-4) is violated then in the set IL defined by (2.2-6) there exists a plant p for which the closed-loop system with the controller ¢ is unstable, However, if the set I was obtained by approximating the true uncertainty regions with disks (see Sec. 2.2.2) and therefore contains plants not present in the original uncertainty set then condition (2.5-4) is generally only sufficient for the original uncertainty set. Nevertheless, the necessity implies that at least for the set II defined by (2.2-6), this is the tightest robust stability condition that can be derived. ‘Theorem 2.5-1 is most instructive for several reasons. First note that ro- bust stability imposes a bound on the oo-norm of the complementary sensitivity function # weighted by Tm. Recall, on the other hand, that for the Hs-optimal control formulation the performance specifications were expressed in terms of the 2-norm of the sensitivity function € weighted by w. The different norms are an indication that performance, robust stability and robust performance cannot be treated “jointly” when the ISE is employed as a performance objective. One of the main motivations for introducing H..-optimal control is that the robust sta- Dility condition and the performance specification are expressed in terms of the same norm. Furthermore recall that measurement noise also tends to impose a bound on the magnitude of 7 (see 2.4-2). However, in process control the constraint imposed by model uncertainty (2.5-4) tends to dominate. Also note that for high frequencies |jc| is small and therefore ‘Thus, for large w (2.5-4) reduces to bel < ay lange (25-6) u CHAPTER 2. FUNDAMENTALS OF SISO FEEDBACK CONTROL ‘The design implication is that the controller gain for high frequencies is limited by uncertainty. The loop gain |pc| has to be shaped to fall below the uncertainty. bound |éa{"? Finally note that in the H,,-optimal control framework we want to minimize |létw]|o for performance (see (2.4-19)) and [fi}&nlloo for robustness. A trade-off between performance and robustness arises from the fact that @ and # are not independent (see (2.4-6)) and making one small will automatically make the other one large. ‘This problem is inherent in feedback control and cannot be removed by clever controller design. The objective of a design method for robust control systems is to reach the best compromise between the conflicting objectives of performance and robustness. 2.6 Robust Performance Robust stability is the minimum requirement a control system has to satisfy to be useful in a practical environment where model uncertainty is an important issue. However, robust stability alone is not enough. If the bound (2.5-4) is satisfied for a family IT, then there exists a particular plant p € II for which the closed-loop system is on the verge of instability and for which the performance is arbitrarily poor. Thus we also have to make sure that some performance specifications are met for all plants in the family II. We will derive conditions for robust performance when the performance is measured in terms of both the 2norm and the oo-norm. 2. 1 Hz Performance Objective Expanding the scope of Sec. 2.4.4 dealing with a single plant p, we are now looking at the family of plants II defined by (2.2-6). We set as our objective to design the controller such that the error resulting from a specific input d is minimized for the “worst” plant in the family. The worst plant is the one giving rise to the largest error. pees sian mina hat = mionaae [alee Wee Ge-9) ‘The integral is maximized by maximizing the integrand at every frequency. From. geometric arguments (Fig. 2.5-1) we find H+ p> |N+Pel bell, — Yp eT @6-2) or Yell (26-3) 26, ROBUST PERFORMANCE 3 where @ = (1+fe)"!, Note that when the robust stability condition (2.5~4) holds the denominator of the RHS of (2.6-3) is positive. Thus (2.6-1) can be rewritten 1 Pre mins [ [lela (26-4) Because the controller ¢ appears in a very complex manner in (2.6-4), there is little hope of finding a simple solution to this problem. However, for a specific controller c, we can use (2.6-3) to determine the worst error which can result from a specific input v for a set of plants II. e- 1 1 Pie, 5 angel = 5 haga eae @6-5) ‘The first factor in the integral accounts for the effect of model error. Because of the robust stability condition (2.5~4) the integrand is always bounded. As expected, the observed error increases relative to the nominal error as the stability boundary is approached. Equation (2.6-5) is a handy analysis tool. After a controller has been designed, performance bound in terms of ISE can be established. The error which is observed on the unknown “real” system as a result of a specific input v is always Jess than the established bound ~ provided the uncertainty description is correct. 2.6.2 Hx Performance Objective If the performance specifications are stated in the Huy framework, then we will require that (2.4-21) be met for the “worst” plant, rg lleu|| 2 = mgxsup [ewiw)] <1 (26-6) [lewlloo = sup lew(iw)| <1 Yp eT (26-7) Using (2.6-3), (2.6-7) can be rewritten as ww (26-8) i (ikal + leu] <1 Wo (26-9) ‘Theorem 2.6-1 (Robust Performance). Assume that all plants in the family oe iP(iw) — Pliw) mm fo: PS s zato)} (2.2-6) 6 CHAPTER 2. FUNDAMENTALS OF SISO FEEDBACK CONTROL have the same number of RHP poles. Then the closed-loop system will meet the performance specification [lew] =supleul<1 pel ‘if and only if the nominal system is closed-loop stable and the sensitivity function Z and the complementary sensitivity function i) satisfy lial + [ew] <1 Yo (26 - 10) ‘rivially, robust performance (2.6-10) implies robust stability (2.5~4) and nominal performance (2.4-20). The interdependence of @ and jj makes it a chal- lenge to meet (2.6-10). Improving the nominal performance (decreasing |éw|) worsens the robustness (increases |jjém|) and pushes the system closer to the point of instability for some plant p € II. If £,(0) <1 then the bound (2.6-10) can be satisfied for low frequencies by making |pc| large; see (2.4-23). If w < 1 for high frequencies the bound (2.6-10) is met by choosing |fc| small; see (2.5-6). By shaping the loop gain |pe| to be large for low frequencies and small for high frequencies the robust performance specifications can be met assuming that they are not too tight (w large) in a frequency range where the uncertainty Z, is large (design by “loop shaping”). A controller which optimizes robust performance solves mijn sup({iifm| + [ew|) (26-11) Compare problem (2.6-11) for optimizing robust performance with problem (2.4~ 19) for optimizing nominal performance. For nominal performance the weighted co-norm of the sensitivity function is to be minimized. For robust performance the co-norm ceases to be a suitable measure. As we will learn in Chap. 11 the appropriate measure for robust performance is the Structured Singular Value (SSV) which will be discussed in the context of multivariable systems. The LHS of (2.6-10) turns out to be the SSV of a particular matrix. We see from (2.6-10) that we can attain robust performance simply by sat- isfying both the robust stability test (2.5~4) and the nominal performance test (2.4-20) with some margin: If |ém| < a(w) and [éw| < 1—a(w) where a < 1,¥w, then robust performance is automatically guaranteed. Thus, for SISO systems, robust performance is not an objective which is particularly difficult to deal with. ‘This will not be true for MIMO systems. 2.7 Summary For the design of a feedback controller the following essentials have to be specified: 2.7. SUMMARY a7 @ process model model uncertainty bounds © type of inputs # performance objectives All design techniques aim to make the error resulting from certain external inputs small. This is equivalent to making the sensitivity function ae 24-1) T+pe ¢ ) ‘imal controller minimizes the ISE for a particular input or equivalently it minimizes the 2-norm (“average”) of ¢ weighted by the input. Hy: mijn [* Pat = min ew (24-14) ‘The H,.-optimal controller minimizes the worst ISE which can result from a set of inputs or equivalently it minimizes the cc-norm (“peak”) of € weighted by w. Hee: ja ew For the suppression of measurement noise and in particular for robust stability the complementary sensitivity ij should be small in the frequency range where the model uncertainty measured by & is large. njn sup few] (24-19) Robust Stability: Hib nlloo up [en] <1 (25-4) Because 1 (24-6) a basic tradeoff between nominal performance and robust stability arises. Robust Performance: libn| + lew] <1 Yur (2.6 — 10) Robust performance is achieved simply by satisfying both nominal performance (éw| <1) and robust stability (Ji@n| < 1) with some margin 38 CHAPTER 2. FUNDAMENTALS OF SISO FEEDBACK CONTROL 2.8 References 2.1. The inverse response behavior of NMP systems is proven, for example, by Holt & Morari (1985a). 2.2.2, The uncertainty description via norm-bounded perturbations (“disks”) was popularized by Doyle & Stein (1981). 2.3. Desoer & Lin (1985) investigated the stability of a large variety of feedback structures in the same manner as in this section. 2.4.2. Horowitz (1963) is generally credited with having first proposed and studied the two-degree-of-freedom controller. 2.4.4, Newton first presented an analytic feedback design technique to mini- mize an Hz objective. The procedure is described in the book by Newton, Gould & Kaiser (1957). 2.4.5. The Heo-optimal control problem was first formulated by Zames (1981). Closely related ideas proposed by other authors are discussed in his paper. A possible limiting response which proves that (2.4-17) is am equality is proposed by Desoer & Vidyasagar (1975). 2.5. The robust stability condition (at least for the MIMO case) is generally attributed to Doyle & Stein (1981). 2.6.2. The robust performance condition follows as a special case from the MIMO Structured Singular Value Analysis proposed by Doyle (1984, 1985) Chapter 3 THE SISO IMC STRUCTURE FOR STABLE SYSTEMS ‘The Internal Model Control (IMC) structure will be introduced as an alternative to the classic feedback structure. Its main advantage is that closed-loop sta- bility is assured simply by choosing a stable IMC controller. Also, closed-loop performance characteristics (like settling time) are related directly to controller parameters, which makes on-line tuning of the IMC controller very convenient. A two-step design procedure will be proposed. In the first step the controller is designed for optimal setpoint tracking (disturbance rejection) without regards for input saturation or model uncertainty. In the second step the controller is detuned for robust performance. While this procedure has no inherent optimal- ity characteristics, it constitutes a simple and transparent approach for finding controllers which satisfy all typical practical requirements. 3.1 IMC Structure ‘The block diagram of the IMC loop is shown in Fig. 3.1-1A. Here p denotes the plant and pm the measurement device transfer functions. In general neither p nor Pm are known exactly but only their nominal models p and jy are available. ‘[Link] function py describes the effect of the disturbance ’ on the process output y. The measurement of y is corrupted by measurement noise n. The controller g determines the value of the input (manipulated variable) u. The control objective is to keep y close to the reference (setpoint) r. Commonly we will use the simplified block diagram in Fig. 3.1-1B. Here d denotes the effect of the disturbance on the output. Exact knowledge of the output y is assumed (py = 1,n = 0). Note that the complete control system to be implemented through computer software or analog hardware is contained in the shaded box in Fig. 3.1-1B. Be- cause in addition to the controller q it includes the plant model p explicitly we 30 0 CHAPTER 3. THE SISO IMC STRUCTURE FOR STABLE SYSTEMS Figure 3.4-1, General (A) and simplified (B) block diagram of Internal Model Control system, ai ik ia as 32. STABILITY CONDITIONS FOR IMC a refer to this feedback configuration as Internal Model Control (IMC). ‘The feedback signal is (p-Buta @a-1) If the model is exact (p = jp) and there are no disturbances (d = 0), then the model output j and the process output y are the same and the feedback signal d is zero. Thus, the control system is open-loop when there is no uncertainty . no model uncertainty and no unknown inputs d. This demonstrates very instructively that for open-loop stable processes feedback is only needed because of uncertainty. If a process and all its inputs are known perfectly, there is no need for feedback control. The feedback signal d expresses the uncertainty about the process. 3.2 Stability Conditions for IMC 3.2.1 Internal Stability In order to test for internal stability we examine the transfer functions between all possible system inputs and outputs. From the block diagram in Fig. 3.2-1 wwe find that there are three independent system inputs and three independent outputs. As shown in Fig. 3.1-1 we choose the independent inputs to ber, ui, and uy and the independent ontputs y,u, and g. If there is no model error (p = ), then the inputs and outputs are related through the following transfer matrix. (*) ( (1-pa)p °}(*) u qa -pq (Ol fun (32-1) o] \pq -eq p/w ‘Theorem 3.2-1 follows trivially by inspection. ‘Theorem 3.2-1. Assume the model is perfect (p = jf). Then the IMC system in Fig. 3.1-1B is internally stable if and only if both the plant p and the controller q are stable, This result is not unexpected. Recall that we pointed out in the last section that the IMC system is effectively open-loop when there is no uncertainty. Be- cause the stabilization of open-loop unstable systems requires feedback the IMC structure cannot be applied in this case. It can be argued that the lack of model uncertainty is an artificial assumption. Uncertainty gives rise to feedback and thus it could be possible to stabilize an unstable system with IMC. However in any practical situation it is unacceptable to rely on model uncertainty for stability. 2 CHAPTER 3. THE SISO IMC STRUCTURE FOR STABLE SYSTEMS Figure 9.9-1. Block diagram for devivation of internal stability conditions. 3.2.2. Relationship with “Classic Feedback” ‘The manipulations necessary to transform the block diagram in Fig. 3.2-2A into the one in Fig. 3.2-2B leave the signals u and y unaffected. If we combine the two blocks g and f in Fig. 3.2-2B, which are both part of the control system, into ‘one block ¢ then we obtain a classic feedback control system with lia as as (32-2) On the other hand if we add the two blocks j to the classic feedback system in Fig. 3.2-2C the signals u and y also remain unaffected. The IMC structure follows with q (32-3) T+ pe ‘Thus, in the way the outputs « and y react to inputs rand d the classic feedback structure and the IMC structure are entirely equivalent and the controllers ¢ and q are related through (3.2-2) and (3.2-3) respectively. Consider specifically the case that p is stable and that p =p. i, Assume that the IMC structure (Fig. 3.2-2A) is internally stable — i.e., that q is stable. Then the equivalent classic feedback structure (Fig. 3.2- 2B) is stable because the internal signals u and y are unaffected by the transformation, ii, Assume that the classic feedback structure (Fig. 3.2-2B) is internally stable. ‘Then (2.3-1) is stable and therefore q defined by (3.2-3) is stable. Thus the ‘equivalent IMC structure is internally stable. ‘We have proven the following theorem. 32. STABILITY CONDITIONS FOR IMC Figure 3.2-2, Alternate representations ofthe IMC structure. “ CHAPTER 3. THE SISO IMC STRUCTURE FOR STABLE SYSTEMS ‘Theorem 3.2-2. Assume that p is stable and that p = p. Then the classic feedback system with controller a (32-2) is internally stable if and only if q is stable. ‘This theorem has very profound implications for control system design. Usu- ally a design procedure involves the search for a controller c such that the closed- loop system has certain desired properties. This search is greatly complicated by the fact that only those controllers c are allowed for which the closed-loop system is stable. Theorem 3.2-2 provides a simple parametrization of all stabilizing con- trollers c for the plant p in terms of the stable transfer function q. Thus, instead of searching for ¢ it is possible to search for q without any loss of generality. ‘This search is much simpler because stability of g automatically guarantees the stability of the closed loop system with ¢ determined from (3.2-2) 3.2.3 Implementation: Classic Feedback Versus IMC In the previous section we showed that IMC provides a convenient parametriza- tion of the classic feedback controller. However, there are significant advantages to not only designing the controller via IMC but also implementing it in the IMC structure as a model block j and a controller block g. In any practical situation u is constrained by some upper and lower bounds (eg., valve saturation). It is well known that these constraints can cause problems for the classic feedback controller. If the controller includes integral action they lead to “reset windup.” Even if there is no integral component, a system which is stable without constraints can be destabilized when constraints become active. In the classic feedback configuration these issues have to be dealt with in an ad-hoc fashion (“anti-reset windup”). If the controller is implemented in the IMC configuration, input constraints do not cause any problems provided the actual (constrained) plant input is sent to the model rather than the input computed by the controller q (Fig. 3.2-3). Then, under the assumption that p = j, the IMC structure remains effectively open-loop and stability is guaranteed by the stability of p and q, as is the case in the absence of constraints. Omitting the saturation block in front of the model block would cause the model to behave differently from the plant when the inputs reach constraints, This model/plant mismatch can lead to instability as we will show later. In the structure in Fig. 3.2-3 there is no information feedback: the feedback signal is always d and the controller q is entirely unaware of the effect of its 33, PERFORMANCE OF IMC 6 Figure 8.2.3. IMC implementation in the presence of actuator constraints actions. In particular, it does not know if and when the manipulated variable saturates. It is reasonable to assume that the closed-loop performance could be improved if the controller adjusted its actions when the manipulated variable has saturated. ‘Thus, while the structure in Fig. 3.2-3 guarantees global stability (Gor p =f), performance might suffer somewhat compared to other anti-windup schemes. It should also be noted that even if g, p, and p are general nonlinear operators the IMC system would still be effectively open-loop (as long as p = p) and the sta- bility of p and g would imply closed-loop stability. ‘Thus the IMC structure offers ‘the opportunity of implementing complex nonlinear control algorithms without generating complex stability issues. 3.3 Performance of IMC 3.3.1 Sensitivity and Complementary Sensitivity Function For IMC (Fig. 3.1-1) we easily find the following transfer functions relating inputs and outputs: (33-1) ‘The sensitivity function ¢(#) relates the external inputs r and d to the error e=y-r ot . d-r d 1+qp-5) We find the complementary sensitivity function (s) by subtracting (3) from unity: Aas) (33-2) e P13 ws) (33-3) 46 CHAPTER 3. THE SISO IMC STRUCTURE FOR STABLE SYSTEMS Figure 3.3-1, Two-degreeof-freedom IMC. When the model is exact (p = j), (3.-2) and (3.3-3) reduce to as) =1-pa (33-4) als) = bq (33-5) We concluded in Chap.2 that ¢(s) determines performance and 7(s) robustness. ‘Through the IMC parametrization the controller q is related to «(s) and (8) in a very simple (linear) manner which will make the design of q easy. The effect of the classic controller ¢ on «(s) and 1(s) (2.4-1 and 2.4-2) is much more complex. 3.3.2 Two-Degree-of-Freedom Controller If both good tracking of r and good disturbance rejection are important and if the dynamic characteristics of the two inputs r and d are substantially different it is advantageous to introduce another controller block (Fig. 3.-1). The effects of r and d on the error are described by 7° ie 2 “TF adlp Tad 63-9) For p = 5 (33-6) becomes e= (I~ paid ~(1~ par 3-7) Here gq is designed for disturbance rejection and q, for setpoint tracking. 3.3.3 Asymptotic Properties of Closed-Loop Response (System Type) “System types” were defined in Sec. 2.4.3 to classify the asymptotic closed-loop behavior. ) | | i 3.3. PERFORMANCE OF IMC ar ‘Type m: sen (8) cs lig =O OSk

You might also like