ADVANCED PROCESS CONTROL Butterworths Boston London Singapore Sydney Toronto WellingtonCopyright © 1989 by Buiterworh Publishers, a division of Reed Publishing (USA) Inc. Allright reserved “This Bunerwortseiton is an unabridged reprint of the book originally published by McGraw-Hill i 1981 No part ofthis publication may be reproduce, stored in a revival system, or tansmited, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior writen permission ofthe publisher Library of Congress Cataloging in-Publiation Data Ray, W. Harmon (Willis Harmon). 1940- ‘Advanced process contol / W. Harmon Ray em. — (Butlerworts series in chemical engineering) Reprint. Originally published: New York: McGrav-Hil 1981, Originally published in series; McGraw-Hill chemical ‘engincering series. Includes bibliographical references. ISBN 0-409-00231-4 1, Chemical process control. 1. Tide. UI. Series ‘TPiss.75.Ra8 1989 s9.23072 664" 2815420 cr British Library Cataloguing in Publication Data Ray, W. Harmon (Wills Haron) 1940- ‘Advanced process conto. 1. Chemical engineering. Process conto. Applications of computer systems. 1 Tide 660.281 ISBN 0-409-90231-4 Buterworth Publishers 80 Monvale Avenue Stoneham, MA 02180, loos76s4321 Printed inthe United States of AmericaThis book is dedicated to my inquisitive students, who surely have been my best teachers.BUTTERWORTHS SERIES IN CHEMICAL ENGINEERING SERIES EDITOR ADVISORY EDITORS HOWARD BRENNER ANDREAS ACRIVOS Massachusetts Institute of Technology The City College of CUNY JAMES E. BAILEY California Institute of Technology MANFRED MORARI California Institute of Technology E, BRUCE NAUMAN Rensselaer Polytechnic Institute ROBERT K, PRUD'HOMME Princeton University NEW TITLES Chemical Process Equipment: Selection and Design Stanley M. Walas Chemical Process Structures and Information Flows Richard S.H. Mah Computational Methods for Process Simulation W. Fred Ramirez Constitutive Equations for Polymer Melts and Solutions Ronald G. Larson Fundamental Process Control David M. Prett and Carlos E. Garcia Gas-Liquid-Solid Fluidization Engineering Liang-Shih Fan Gas Separation by Adsorption Processes Ralph T. Yang Granular Filtration of Aerosols and Hydrosols Chi Tien Heterogeneous Reactor Design Hong. H. Lee Molecular Thermodynamics of Nonideal Fluids Lloyd L. Lee Phase Equilibria in Chemical Engineering Stanley M, Walas Physicochemical Hydrodynamics: An Introduction Ronald F. Probstein ‘Transport Processes in Chemically Reacting Flow Systems Daniel E. Rosner Viscous Flows: The Practical Use of Theory Stuart W. Churchill REPRINT TITLES Advanced Process Control W. Harmon Ray Elementary Chemical Reactor Analysis Rutherford Aris Reaction Kinetics for Chemical Engineers Stanley M. WalasCONTENTS Preface Introduction ‘Why Advanced Process Control What Is “Modern” Control Theory? Mathematical Models for Process Dynamics Introduction to On-Line Data Acquisition and Computer Control Introduction Computer System Architecture ata Acquisition and Control Some Examples Control of Lumped Parameter Systems Introduction Linear Multivariable Control Systems Optimal Control Theory and Practice Nonlinear Multivariable Control Discrete Time Systems Control of Distributed Parameter Systems Introduction Feedback Control of Linear Distributed Parameter Systems Optimal Controt Theory and Practice Feedback Controller Design for Nonlinear Distributed Parameter Systems Control of Systems Having Time DelaysChapter 5 st 52 53 sa 3s 36 37 Chapter 6 61 62 63 65 66 State Estimation and Stochastic Control Introduction State Estimation for Linear Systems Described by Ordinary Differential Equations State Estimation for Nonlinear Systems Described by Ordinary Differential Equations Stochastic Feedback Control for Systems Described by Ordinary Differential Equations ‘State Estimation for Systems Described by First-Order Hyperbolic Partial Differential Equations ‘State Estimation for Systems Described by Second-Order Partial Differential Equations ‘Stochastic Feedback Control for Distributed Parameter Systems Control System Design Case Studies Introduction Control of Mult-sidestream Distillation Columa ‘The Control of a Multiple Effect Evaporator [A Strategy for Steel Mill Soaking Pit Control Control of Metallurgical Casting Operations Further Case Studios Appendix: Some Computer-Aided Design Programs Indexes Author Index Subject Index 245 4s 2 mm 28 23PREFACE This book is designed to be used as a text for advanced undergraduate and graduate courses in process control as well as to serve as a reference for the practicing control engineer. Bearing in mind the likely background of the reader, the mathematical prerequisites consist of introductory undergraduate process control material, elementary matrix algebra, and an introduction to differential equations. The standard undergraduate chemical engineering curriculum usually provides the necessary background. In spite of these modest mathematical demands, the book strives to provide a broad coverage of applied modern control theory. Although only one or two sections may be devoted to theoretical subjects upon which entire books have been written, the goal is to provide the process control engineer a synopsis of the theory, a clear exposition of the most useful design techniques, and example problems to demonstrate the key features of the computational algorithms. Numerous references are provided so that the interested reader may delve more deeply into the mathematical theory or seek further illustrative examples. ‘The material has been tested in the classroom over the last five years both at the State University of New York at Buffalo and at the University of Wisconsin. ‘While the bulk of the students have been undergraduate and graduate students in chemical engineering, these classes have also included both electrical and ‘mechanical engineering students interested in process control. Usually these courses have involved laboratory projects through which students are able to implement some of their control system designs. It is perhaps useful to point out the thrust of each chapter. In Chapter 1 there is a discussion of the engineering and economic incentives for implementa- tion of advanced process control concepts. The key elements of an advanced control scheme are discussed and their integration into a coherent, compre- hensive structure is outlined. This is followed by an overview of the different types of process dynamics models normally encountered. This provides a guidesil PREFACE for the ensuing chapters, which treat control system design for each type of process model. ‘Chapter 2 is an introduction to the computer and interfacing technology needed for real-time process control. The basic structure of microcomputers and ‘minicomputers is discussed together with the features of peripheral devices. Both digital and analog data acquisition interfaces are covered, and their essential features are discussed. A substantial number of process sensors, control actua- tors, and transducers are described in the context of their typical applications. Finally, several actual process control interfacing projects are presented and the ultimate design strategy with regard to sensor selection, multiplexing, signal conditioning, mode of transmission, etc, is discussed. This chapter is designed to provide the reader a good grasp of the considerations necessary to implement computer control schemes. The intention is to give the needed perspective and ‘motivation for the more theoretical material to follow. ‘Chapter 3 deals with the problem of control system synthesis for lumped parameter systems. Both differential equation state space models and transfer function models are considered, and the equivalence of the two formulations is demonstrated. Standard design procedures for linear multivariable systems are introduced to demonstrate that these “suboptimal” procedures often produce hhigh quality, but simple, designs. Then classical optimal control theory is presented, both to provide a basis for comparison with simpler suboptimal strategies and to lay a theoretical foundation which can be built upon in later discussions of state estimation and stochastic control. Finally, several methods ‘of nonlinear system controller design are presented. Chapter 4 introduces the reader to distributed parameter systems which are ‘modeled by partial differential or differential-delay equations. As in the previous chapter, there is a discussion of the types of distributed systems commonly encountered in the process industries and the appropriate techniques for control system design. Following this, optimal control theory is presented and optimal controller design illustrated by examples. Where applicable, modal decomposi- tion for linear systems and Galerkin procedures for nonlinear systems are stressed. The chapter ends with a treatment of systems having time delays. The goal of this chapter is to demonstrate that through appropriate analysis, distrib- uted systems are not significantly more difficult to handle than lumped systems. Chapter 5 is devoted to state estimation and stochastic control. Although the relevant statistical theory can get rather deep at times, particularly for nonlinear systems, the essential features are presented in an operational-intuitive manner suitable for the reader with only a rudimentary background in statistics. This approach is designed to allow the control engineer to see the thrust of the analysis and understand the important results so that useful estimation and control algorithms may be developed. Chapter 6 consists of a series of detailed case studies of control system design. For each control problem treated, the reader is led through the steps of control system synthesis and then shown the performance of the resulting design when applied to a pilot plant process. It is thought that these exercises inPREFACE xl .gnthesis, which draw upon all the techniques covered in the earlier chapters, will demonstrate the power of these sophisticated and useful process control ‘methods in a practical context. ‘There are many who contributed their efforts to this book. Several genera- tions of students provided example problems, criticism, and testing of the homework assignments; Klavs Jensen and Tunde Ogunnaike were particularly helpful in carrying out computations for example problems. Proofreading help was provided by Klavs Jensen, Tunde Ogunnaike, Joe Schork, and Marvin Schwedock, while Alan Schmidt kindly served as photographer for some of the half-tones. Professors Jay Bailey, Ram Lavie, Manfred Morari, John Seinfeld, and Joe Wright read the manuscript and contributed many helpful suggestions. The detailed comments resulting from the classroom testing of the material by my colleague, Manfred Morari, were especially valuable. Finally there were the typing efforts of Diane Petersen and my wife, Nell. Their skill transformed barely legible scribbling into flawlessly typewritten copy, while their patience allowed numerous revisions of the original manuscript. As usual, this project could not have been completed without the devotion of my wife, Nell, who participated in all aspects of the book (typing, preparation of the index, proofreading, etc.) while keeping up with all her other activites. The author is sreatly indebted to all who contributed to this venture. W. Harmon RayADVANCED PROCESS CONTROL‘CHAPTER, ONE INTRODUCTION 1.1 WHY ADVANCED PROCESS CONTROL From its beginnings in antiquity [1, 2] until the early 1960s, the field of process control was based almost entirely on mechanical, electrical, or pneumatic analog controllers, which were usually designed using linear single-input single-output considerations. Hardware limitations, economic cost, and the dearth of applica- ble theory usually precluded anything more complex than these simple schemes. Because many large-scale industrial processes are endowed by nature with large time constants, open-loop stability, and significant damping of fluctuations through mixing and storage tanks, such simple control schemes work well for pethaps 80 percent of the control loops one might encounter. For the remaining 20 percent more difficult control problems, most controllers were considered marginally acceptable during this early period because there were few environ- mental regulations, product specifications were quite loose, and intermediate blending tanks could cover many of the sins of inadequate control. Thus the costs of even small sophistications in control were high and the economic incentives for improved control were comparatively low. Over the last 10 to 15 years, there has been a dramatic change in these factors. Industrial processes are now predominantly continuous with large throughputs, highly integrated with respect to energy and material flows, con- strained very tightly by high-performance process specifications, and under intense governmental safety and environmental emission regulations. All these2 ADVANCED PROCESS CONTROL 9 THT TH B80 Yeu Figure L.1An example of price trends for realtime minicomputers, factors combine to produce more difficult process control problems as well as the requirement for better controller performance. Significant time periods with off-specification product, excessive environmental emissions, or process shut- down due to control system failure can have catastrophic economic con- sequences because of the enormous economic multipliers characteristic of high- throughput continuous processes. This produces large economic incentives for reliable, high-quality control systems in modern industrial plants. ‘Another recent development in process control is that the performance of real-time digital computers suitable for on-line control has improved signif cantly, while prices have fallen drastically. Figure 1.1 shows an example of the price trends for small minicomputers in spite of the inclusion of more reliable electronics and increasing inflation, With the process control computer now such a small part of the overall process capital costs, the installation of a fast minicomputer with large amounts of storage can often be easily justified on the basis of improved safety and manpower savings. Once in place, the computer is usually operating in a timesharing mode with large numbers of input/output operations, so that the central processing unit (CPU) is typically in use only about 5 percent of the time. Thus many installations have 95 percent of the computing power of a highly capable minicomputer, programmable in a high- level language such as Fortran, already available for implementing sophisticated computer control schemes. ‘At the same time, modern control theory has been under intense develop- ‘ment, with many successful applications in the aerospace and aircraft industries. Most recently, a number of process control research groups have been applying these ideas to simulated, laboratory, and even a few full-scale plant processes, so that much of the applicable theory needed for sophisticated process control algorithms is at hand. However, there remains the problem of communicating, these results, in a readily applicable form, to the process control engineer who ‘must design an economically optimal process control scheme. The present text is directed toward this educational need, and strives to present a comprehensive introduction to the theory and practice of modern computer process control.srRoDUCTION 3 1.2 WHAT IS “MODERN” CONTROL THEORY? In contrast to classical control theory, which is essentially limited to single-input single-output systems described by linear differential equations with constant coefficients (or their corresponding Laplace transforms), so-called modern con- trol theory has developed to the point where results are available for a wide range of general multivariable systems, including those described by: 1, Linear, variable-coefficient differential equations 2. Nonlinear differential equations 3. Differential-difference and other hereditary equations 4, Partial differential and integral equations ‘The results of modern control theory include so-called optimal control theory, which allows the design of control schemes which are optimal in the sense that the controller performance minimizes some specified cost functional. In addition to controller design, modern control theory includes methods for process identification and state estimation. Process identification algorithms have been developed which determine the model structure and estimate the model parameters, either off-line or adaptively on-line. These are useful both in the initial control system design and in the design of adaptive control systems which respond to such changes in the process characteristics as might arise, for example, with the fouling of heating exchanger surfaces or the deactivation of catalyst in chemical reactors. State estimation techniques are on-line methods of ‘estimating those system state variables which are not measured and for improv- ing the quality of all the state-variable estimates when there are measurement ‘errors. In processes where numerous sensors are not available or are expensive, ‘on-line state estimation can be of significant practical importance. It is perhaps useful to demonstrate how all these components of a compre- hensive computer process control scheme might fit together for a particular process and to point to where the technical details are discussed in the text. Figure 1.2 outlines such a control scheme which consists of the following: 1. Process which responds to control inputs u; natural process disturbances dy; and special input disturbances d, which enhance identification. The true Process state x is produced, but this is seldom measured completely or without error so that one employs 2. Measurement devices, which usually are able to measure only a few of the states or some combination of states and are always subject to measurement error. These are discussed in Chap. 2. The measurement device outputs y are fed to the 3, State estimator, which makes use of the noisy measurements y along with a process model to reconstruct the best possible process state estimates &. The state estimator must have parameters P which are calculated (either off-line= | eee | oe »} ccototer bet sue | Se] ie 4 | 7! Sie | oe | tt etne lef | caste | | = t = - mop ns Biel [me J Figure 12 A comprehensive computer control scheme, ‘or periodically updated on-line) by algorithms, possibly requiring the measurement outputs y and state estimates & as well as the process model parameters a. The details of state estimation are found in Chap. 5. The process state estimates are passed to the 4. Controller, which calculates what control actions must be taken based on the state estimates %, the set points uy, x, (which themselves may be the subject of process optimization (3)), and the controller parameters K. The controller parameters K can be calculated either off-line or adaptively on-line based on current estimates of the model parameters @ and process state &. Techniques for controler designs are discussed in Chap. 3 and 4. The process model parameters must be determined from theINTRODUCTION 5, 5. Process identification scheme, which takes measurements from the process as raw data y (and may choose to introduce experimentally designed input disturbances d,) in order to identify the process model parameters a. If the parameters are invariant with time, then this need be done only once; however, if the process changes: with time, then the process identification scheme may be activated periodically to provide adaptation to changing conditions. Process identification is not treated in great detail in this book because there are many fine existing reference books in this area [4~10}, In most applications only a few of the components of this control structure are required. Chapter 6 describes some control system design case histories which illustrate this point. 1.3 MATHEMATICAL MODELS FOR PROCESS DYNAMICS For all but the simplest control schemes to be effective, some description of the process to be controlled must be available. Usually this description is a mathe- ‘matical model. In classical single-loop process control theory, this model often takes the form F(s) = 8(s)ii(s) + 84(s)d(s) (13.1) where d(s), i(s), 7(s) are the Laplace transforms of known disturbances d(1), the controller input u(), and process output y(1), respectively. Here g(s), 8) are the scalar transfer functions relating them as shown in Fig. 1.3. However, many processes are much more complicated and require more complex models. We shall discuss some of the naturally occurring models in what follows. Linear Multivariable, “Lumped” Models “Lumped parameter” or “lumped” models refer to process descriptions in which there is no spatial dependence; either ordinary differential equation or Laplace transform models are employed. The simplest case of a lumped multivariable process model would be a generalization of the single-input single-output transfer function model (Eq. (1.3.1)] to multiple inputs and multiple outputs, as shown in Fig. 14. In this case there are k disturbances, m inputs, and / outputs Proce Disurtance—el ytd Ay laps] output Flame 13 Singleinput single-output a * la) —inear system.6 ADVANCED PROCESS CONTROL + ys 7 (can) aa Figure 14 Multivariable linear sytem with k disturbances, m inputs, and ¢ outputs, related by H(s) = G(s)(s) + GAs) (1.3.2) where &(), (0), (6) are vectors and G(s) is an / x m matrix, while G(s) is an P< k matrix, Hs) a) HO) =|) ay 2) = FO! aaa) 4 a(s) 70) Sul) 8:2(s) G40) =| PB / i‘ inf) sao) £0) Bld) > als) ete) a 2809 Fol Bal sl) sels) su) ‘An alternative form of linear, lumped parameter models is the time-domain ‘model: FO o Axi + Bult) + Tas) x(0) = xy (3a) y= exe) (135) where the variables (0), i= 1,2,...,m, are the state variables; the »4(1), 4m 1,2,...44, are the output variables; the u,(0), r= 1, 2,...m, are the pu ‘ariables or control variables; and the di), $= 1,2, .., k, are the disturbance variables,eTRODUCTION 7 ‘The form of the ‘ransform-domain model [Eq. (1.3.2)] or time-domain model (Eqs. (1.3.4), (1.3.5)] may be chosen by convenience because they may be shown to be equivalent if A, B, C, and D are constant matrices. The basis of these ‘multivariable models may be determined from the fundamental differential equations of physics or, more often, from empirical fitting of the model equa tions to measured process dynamics. In either case, there is a wide range of processes which may be approximately modeled in this form. More will be said about these models in Chap. 3. Nonlinear Lumped Models Systems described by nonlinear ordinary differential equations are very common in the process industries and have models of the general form SO - t(x(0,(),0) x0) = x (136) ¥(0) = B(x(0), (0) 37) where as before x(1) represents a vector of states, y() the outputs, do) the disturbances, and ui) the control inputs. The nonlinear functions f and ly may have a variety of forms and are usually derived from the momentum, energy, and material balances and the type of sensors used for the process under study. ‘As an example, a liquid-phase reaction, A> B, carried out in @ continuous stirred tank reactor will have a nonlinear model of the form ae * Giw a+ Dall ~ dere 7) + (138) ae, % SB = — (1+ A), + Dall — seo 275) + pu +d; (39) where x, is the conversion of A to product and x, is a dimensionless reactor temperature. The quantities Da, y, B, and B are parameters, and u is the jacket coolant temperature which is manipulated to achieve temperature control. The variables d, and d, are disturbances in the feed reactant concentration and feed ‘temperature, respectively. The output equation is yre (1.3.10) which indicates that only reactor temperature is measured. The control of this class of processes is discussed in Chap. 3. Distributed Parameter Models A significant number of industrial processes are “distributed” in space so that their behavior depends on spatial position as well as time. These models usually take the form of partial differential equations and are often derived from the8 ADVANCED PROCESS CONTROL, fundamental balances of momentum, energy, and material for the process at hand, As an example, many heat conduction, fluid flow, and chemical reactor processes take the form Batu vn Vs...) (a3) where both the state variables x and control variables u may depend on time {and on a multidimensional spatial positon. A simple illustration ofthis type of distributed model may be found by considering that the liquid-phase reaction ‘AB noted above is now carried out in a tubular reactor with cooling at the wall, A reasonable model of such a reactor is inthe form axy 112,83) ~ Ut Ben OA or, + Dal ~ xese( 7275) (13.2) ax,(r, 2, 1) ax, 11 a/ ax a 0B + tee) + BDa(I - aye 275) (13.13) with boundary conditions 2-0; =a =0 ax, r=0: a 0 (1.3.14) ax, ax, rat Seo Bai x) Here x, and x, are reactor conversion and temperature, respectively, and w isthe wall coolant temperature which can be manipulated. The quantities Peyy Pey. Da, B, f, 7, and Bi, are parameters, and the output equations are of the form WO) = rn tn) 12... (1315) and represent the thermocouples placed at positions (r,s) in the reactor. Note that the control variable in this example problem appears in the boundary condition, Eqs. (1.3.14), as often happens in distributed parameter control problems. Chapter 4 will treat the control of this class of processes in some detail. REFERENCES 1, Fuller, A. T°. Dynamic Systems, Measurement, and Control” Trans. ASME 986: 109,224 ae). 2. Mayr, O: Origins of Fedback Conirol, MIT Press, Cambridge, Mast, 1970,INTRODUCTION 9 23, Ray, W. H, and J. Seekely: Process Optimization, Wiley, New York, 1973. 44 Sage, A. Py and J. L. Mels: System Idenifcaion, Academic Press, New York, 1971 5. Box, G. EP, and G. M. Jenkins: Time Series Anabyus, Holden-Day, San Francisco, 1970. 6 Seinfeld, J H, and L. Lapidus: Mathematical Methods in Chemical Engineering, vol. 3: Process ‘Modeling, Esimation, and Identification, Prentice-Hall, Englewood Cis, NJ, 1974 7, Eykbolf, P.: Sytem Identification, Wiley, New York, 1974, a 9, Ray, W. H, and D. G, Lainiotis; Distributed Parameter Systems, Marcel Dekker, New York, i978 10, Stewart, W. E, and J.P. Sorensen: Collocation and Parameter Extimation in Chemical Reaction Enginerng, Wiley, New York, in press.‘CHAPTER TWO INTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL 2.1 INTRODUCTION In order to effectively design a modern process control scheme, one must have a basic understanding of the practical means of implementation. For example, computational algorithms suitable for on-line application must be carried out faster than real time on a typical process control minicomputer. Thus sampling, frequency, computer computation times, and other factors must be considered in order to design a workable system. The successful installation of a modern control scheme also involves considerations of data links and computer com- ‘mand transmissions. Process variables must be measured and these data pro- vided to the computer, while desired control actions must be computed and transmitted to the process. Figure 2.1 gives the simplified structure of such a computer control scheme. The measured variables are transduced, filtered, and sent by cable to a computer interface for data inputs. This measurement information is then incorporated into the computer control algorithm and control signal outputs generated. These signals are transmitted to local trans- ducers and local controllers (such as valves and heaters), which cause changes in the manipulated variables of the process. The technology which accomplishes these tasks must be generally understood by the control engineer so that the designs are realistic and effective. Thus this chapter shall be devoted to provid- ing the reader an overview of the considerations necessary in actually imple- ‘menting a computer control scheme. 0INTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL 11 Conor Date Proce tut | out af Sone | inerce ‘mrace Somputer Sena ata ie Sieal Proce to Loca Satiionng [*poccy—] teconvatd Pace] avers ba compat inputs ‘igure 2.1 Structure ofa computer proces contol scheme. We shall begin by describing typical minicomputer architecture, capabilities, and peripherals, This should give one an idea of the computer resources which might be typically present in an industrial computer control installation. Sec- ondly, we shall discuss the types of interfaces normally used for data acquisition and control-signal generation. These interfaces put constraints and specifications ‘on the types of signals needed from the measuring devices. Finally, we consider ‘commonly used sensing devices as well as techniques for transducing, multiplex- ing, and conditioning the data and control signals. This latter material should provide the control system designer with a grasp of the quality of the signals, which must be dealt with in design, and can guide the choice of signal mode (eg, analog voltage, analog current, or digital signal), While this chapter is important in providing a well-rounded perspective for the reader, it is not crucial to the theory which follows; thus it is possible, if desired, to bypass this material for the moment and return to it later. 2.2 COMPUTER SYSTEM ARCHITECTURE, There are many different manufacturers of minicomputers designed for process control applications (a few of these are listed in Table 2.1) and many different types of architecture, However, the functions of the computer itself are much the same, and peripherals have been standardized to be compatible for most minicomputer mainframes. Thus it is possible to describe generally a typical minicomputer, We choose as our example the minicomputer in the Department ‘of Chemical Engineering at the University of Wisconsin. This system, shown in Figs. 2.2 and 2.3, consists of a PDP 11/55 minicomputer with associated Peripherals, interfaces, etc, We shall briefly describe each component of the system,12. ADVANCED PROCESS CONTROL Figure 22 Chemical Engineering Minicomputer System, Univesity of Wisconsin, 1 Central processing unit (CPU)—This is the heart of the minicomputer and has overall control of the system. The computer uses the binary digits (0 or 1) as the basic computational and communication code. This is accomplished electronically by manipulating and transmitting pulses which need be inter- preted only as a0 or a 1. The resolution of a machine is dependent on the number of binary digits (abbreviated bits) in the basic unit of information, called the word. Common minicomputers have 8, 12-, 16-, or 18-bit words (see Table 2.1). Usually a smaller unit, a byte consisting of 8 bits, is used as well, because a byte suffices to specify characters (alphabet letters, numbers, control characters, etc.) according to industry standards (e.g, ASCII Code). Figure 23 PDP 11/55 mini- computer at the University of ‘Wisconsin.SIRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL. 13 ‘Table 2.1. Some of the commonly encountered minicomputers for real-time applications* ‘Speed: fixed point ‘Maximum addition, memory Approximate Manufacturer Computer (word length) memory (words) cyele time (us) price Data General 1, Nova (I6 bit) 2K 07,07 Siok Corp. 2. Eclipse ,C (16s) 128-5I2K 06,07 Si0K Digital Equip 1. PDP R(I2bis) = 128K 3015 Siok ment Corp. 2 2K 35.12 S4K 3 128K 20,07 39K 4 sk 03,03, SOK Hewlet- HEP 1000 (16 bits) om K 09,06 Siok Packard Corp. + Current and comprehensive listings of available minicomputers may be found in wade journals, surveys te. (references (9,10). ¢ Price includes CPU and 32 K words of memory Normally 7 bits are used to specify the character and one is used for error checking (parity). Thus 2” = 128 characters may be defined by a byte. ‘Normally the bits in a word will be arranged in octal form by grouping them in3s eg, HO 010100, oot 101 Since 3 bits allows one to obtain the numbers 0 to 7, as shown in Table 2.2, octal integer notation is a convenient shorthand for addresses, machine instructions, etc. For example, the 15 bits shown above represent the octal number 62415, which is a much more efficient notation than the original binary form. A 16-bit machine (the most common type) thus can use | bit for sign (+ or —) and 15 bits to represent integer numbers. The maximum integer range then is from ~77777 to +77777 octal, which corresponds to 32,767 to +32,768 decimal. These octal numbers are sometimes used to enter elementary computer instructions via switches on the CPU console (Fig. 2.4 shows a photograph of a typical console). Associated with the CPU, and ‘working in parallel, many computers have a hardware floating-point processor ‘which performs all floating-point arithmetic operations. For the PDP 11/35 machine shown in Figs. 22 and 23, which has 16-bit words, these floating- point operations involve numbers whose length is two words (32 bits). If double precision is called for, the floating-point processor uses 4 words (64 bits); however, double-precision operations require substantially more time to ‘execute, The precision of an arithmetic operation depends on the number of bits carried along. For a 32-bit single-precision operation on the PDP 11/55, 24 bits are used for the fractional number and 8 bits for the exponent. In each case 1 bit is reserved for the sign, so that the exponent is limited to floating-point numbers between 2~* and 2” (approximately 10-™* to 10°14 ADVANCED PROCESS CONTROL ‘Table 2.2 Binary, octal, and decimal conversion Binary Octal Decimal ot 101, oot 110 ool 111 ‘010 000 ‘ont 000 100 000 101 000 10 000 111 000 (ctal—Binary conversion For 3 its oxtal = by +2, + Ab, where by by, by are binary dite. Decimal-octal concersion Decimal = ay + 84, + 642, $5 Bay, where ay a... ay ar the octal digits decimal). The fractional part, of 24 bits, has an error corresponding ap- proximately to the least significant bit. With I bit for the sign and 23 for the fractional number, the relative error is approximately 2~® = 10-7, so that a maximum of about 6 decimal digits are reliable in single-precision operations. For double precision, 8 bits are used for the exponent and 56 bits for the fraction, giving an approximate relative error of 2~* = 10-". This corre- sponds to a maximum reliability of about 16 decimal digits. For the floating- point processor used on the PDP 11/35, these operations are quite fast (~ 1 uss for addition, ~ 4 us for division). Such high-performance floating-point processors, coupled with logic which allows operation in parallel with the CPU, make modern, inexpensive minicomputers rather fast in carrying out ‘numerical calculations. 2. Memory—The memory devices associated with minicomputers may be of several different types (see Table 2.3). Core memory, which uses electromag- netic storage techniques, has a typical cycle time of ~ 1 yss—rather slow in the present context. However, because of its ferromagnetic character, core memory retains stored information even when the electric power fails. ‘Metal-oxide silicon (MOS) memory, on the other hand, consists of simpleINTRODUCTION 10 ON-LINE DATA ACQUISITION AND COMPUTER CONTROL. 15, Figure 24 The PDP 11/35 CPU operator console (Reproduced by permission of Digital Equipment Corporation). circuit semiconductor devices, which are both cheaper and faster (~ 500 ns cycle time) than core memory. Bipolar transistor memory makes use of more complex integrated circuits which are still faster (~ 300 n s eycle time), Both of these devices, however, lose the stored information in a power failure; thus battery power backup is often available. Special-purpose memory such as read only memory (ROM) is usually used to permanently store basic instruc- tions for starting up the machine, input/output commands, etc., or for crucial, frequently used data files. These devices require special programming hardware in order to make any changes in the stored information. Memory can be purchased with and without extra parity bits. These extra bits are used to double check each byte transfer operation for transmission errors. This feature is useful in crucial computations and in diagnosing the specific locations of any memory failures. For the PDP 11/55 system shown in Figs. 22 and 2.3, all four types of memory devices are in use. Although only 64 K bytes (32 K words) of memory can be routinely addressed by a 16-bit machine, the PDP 11 series can make use of as much as 128 K words of memory by using many memory partitions, each of which may be as much as 32 K words in size, 3. Mass storage devices—There are a number of different types of mass storage devices available for minicomputers (see Table 2.3). These are used for inexpensive high-capacity storage with access times and transfer rates much slower than memory. Magnetic tape is the cheapest bulk storage device, but has the disadvantage that it has very slow access times (due to the time16 ADVANCED PROCESS CONTROL, Table 23 Capabilities of storage devices* Typical Typical ‘Approx. ‘Typical capacity/ access ransferrates cost, Device 6bit words times (words /s)—_$/1000 words Memory: (eyete time) Core 128k Cans Les 50-100 Mos: 8K O2ps OS 50-100 Bipolar 128K Olps O03 ps 500-1000 Mass storage: Fised-head ‘ike ~1 milion ~n 20-0 Moving-head disks 1-100 milion 35-708 108 05-2 Floppy disks 256-512.K 500m 3x 10¢ 5-10 Magnetic Several 2x 1010 ‘ape 10-1Smillion seconds 4x 10 ' * Figures are for 1980, required to move back and forth on the tape drive to find the desired files). For sequentially stored files, magnetic tape has reasonably fast transfer rates and is thus suitable for backup storage of large programs, data banks, etc. A much faster mass storage device is the rotating disk, which comes in several types (fixed-head, moving-head, and floppy). Most of the disks have remov- able packs or cartridges so that interchangeable data files, program libraries, ctc,, are possible. Fixed-head disks usually have a small capacity but short access times and high transfer rates. These devices are used for fast swapping of files between memory and disk. Moving-head disks have much slower access times but transfer rates comparable to those of fixed-head disks and much higher capacity. Floppy disks are low-capital-cost, small-capacity, low- performance mass storage devices in common use with microcomputers. Figure 2.5 shows a photograph of the magnetic tape unit, the movable-head disk and drive, and the floppy-disk drive used with the PDP 11 systems at the University of Wisconsin. 4. Real-time clock—The scheduler and timekeeper of the computer system is the real-time clock. This allows the computer to schedule tasks to be per- formed on a regular basis, such as taking data or changing control variables at specified time intervals, and timesharing the CPU's attention between various terminals and other peripherals. Each time the computer is powered up, the correct date and time are entered, and the real-time clock keeps track of chronological time after this initialization. A different means of sharing the CPU's attention is through interrupts. Each task running on the computer has a certain priority level; when it needs the attention of the CPU (for example, to record a data point), the task generates an interrupt, and if this task has a higher priority than the task currently being serviced by the CPU, the newINTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL. 17 Figure 255 Mass storage devices used on LSI II minicomputer. Floppy disks plus dik drive,18 ADVANCED PROCESS CONTROL task is immediately serviced. Thus one may establish a hierarchy of priorities and interrupts in the computer and have the most urgent tasks (such as data taking) serviced immediately while those less pressing (such as printing on the line printer) must wait. Typical waiting times are a fraction of a second, so that the user is not usually aware of waiting in a queue. 5. Input / ouput interfaces for daia—These interfacing devices deal with either digital or analog signals. Digital 1/0 interfaces are usually capable of either (@) parallel two-way simultaneous transmission (for example, 16 bits input and I6 bits output for a 16-bit machine) at speeds comparable to those within the computer system, or (b) one-way serial transmission (16 bits in only one direction at once for a 16-bit machine) at specified transmission rates. Parallel-transmission interfaces are typically used for high-speed communica tion between minicomputers, to multiplexers, or to other high-speed digital logie devices. Serial-iransmission interfaces are used for communication with slower digital devices, such as terminals or graphical display units. Serial- transmission rates are indicated as baud rates, where baud rate = 10x (# characters/s). Typical rates range from 30 to 960 characters/s, depending con the device being interfaced (see Table 2.4). Both types of digital interface devices can be put under program control and the information analyzed by individual bits, bytes, or words, depending on the application. Table 2.5 lists some typical applications of digital I/O interfaces. ‘Analog 1/0 interfaces involve analog-to-digital (A/D) and digital-to- ‘analog (D/A) conversion. A/D converters convert an analog voltage signal lying within a specific range (such as +10 V, 0 to 5 V, etc.) to an integer number. The maximum size of the integer number, and thus the resolution of the conversion, depends on the number of bits handled by the converter. The range of the integer /is given bY Ing, = 2%, where N is the number of bits, A I2-bit converter with an input Vollage range of ~10V.< V < +10 V differential is shown in Fig. 2.6. The converter handles a differential signal because the reference voltage is also measured. A single-ended input would have only one voltage converted referenced to the computer ground voltage. The maximum integer range for the 12-bit converter is 2'* = 4096, This Table 2.4 Some communications peripherals Typical Peripheral, (Communication rates capital cost, Line printers 150 characters /s1200ines/mia _$3000-$20,000 ‘Typewriter terminals 10-30 characters/s 1000-4000 Video terminals 240-960 charactes/s $1000-$10,000 ‘Storage scope terminals 480-960 characters/s '$3000-$8000 Punched cards 250-1200 eards/min '$3000-$10,000 Paper tape 10-300 charactrs/s '$3000-$5000 Magnetic pe 10-10 characters/e 5000-$15,000INTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL 19 ‘Table 2.5 Some typical applications of data I/O interfaces “Type of interface ‘Typical connections and data Digital inputs Digital outputs Analog inputs ‘Analog outputs From other computers Digital instrumentation ‘Communication terminals Switch settings Relay postions Multiplexer status ‘Alarms Counters Logie devices To other computers Commands to relays and solenoid Commands to switches ‘Alarm messages Stepping motors Sequencing of ‘multiplexers Logic devices (Commands to digital Communications terminals Hard copy and graphical ‘output devices ‘Various control signals From analog computers ‘Analog instrumentation: “Temperatures Pressures Flows Composition ‘To analog computers Graphical ploters Oxciloscopes Process parameter changes Control signals: Heater contols ‘Se-point changes Valve divers number of integers is chosen to run between —2048 < I < + 2047, where 1 = 0 falls between the negative and positive integers. For the positive range (2047), the resolution of the converter is the voltage range divided by the intervals between integers, or Resolution = 12¥ = 0.00489 V lov 2047 This gives an expected error of +} the resolution, or 0.00244 V. In this case20 ADVANCED PROCESS CONTROL the relative error in conversion is approximately given by 0.00244 V. Relative ertor = red vole so that more precision is obtained if the measured voltages are close to the full-scale voltage (10 V in this example). This can be accomplished through signal conditioning, which will be discussed in the next section. The through- put speeds of A/D converters are routinely 50,000-100,000 conversions/s, and high-performance models can achieve even higher rates. As an illustra- tion of A/D conversion, suppose the 12-bit converter shown in Fig. 2.6 ‘measures a voltage and reports the integer 1261 as a result of the conversion. To determine the actual measured voltage, one uses the formula 107, 12,610 ‘meaneed 2047 ~ 2047 where for this input voltage, the relative error is. = 6.16023 V = +0.000397 = +0.0397% a rather small value. D/A converters operate as just the reverse of A/D conversion. An integer with range 2" is placed in a register and converted to an analog voltage output. The 12-bit D/A converter shown in Fig. 2.6 has a voltage output range of -10 V< Vs + 10 V, corresponding to input integers ranging from ~2048 < J’ < +2047. Thus the resolution and expected error of this D/A converter are the same as those of the A/D converter discussed above, To demonstrate how one programs the converter to obtain a desired voltage output, let us assume we wish to output 3.5 V. To choose the input Hise amie fy, | AD LL tes “10 yo V< +10 vote (a) Analog t-tigialconenion (12 bit) ss tegen] A | Ty ntog vote To 2048 12067 sveresoy Figure 2.6 12-bit A/D and D/A (0) Digulsoanalg conversion (12 1) ‘conversion.integer, one uses the formula 2047 Vowipur _ (2047)(3.5) =e ix 716.45 which is rounded {0 jap: = 716. This gives an actual output voltage of 716 Youu ™ Fgh * 10 = 3.49780 very close to the desired voltage. Throughput rates for D/A converters are also very high, typically on the order of a million conversions/s ‘Most computer installations will have multiple channels of A/D and D/A converters in order to allow several different projects to interface with the computer at any one time. For the computer system at the University of Wisconsin, there are 16 channels of 12-bit differential A/D and 16 channels of 12-bit D/A. 6. Communications peripherals—There are many different types of devices for communicating between the user and the computer system (see Table 2.4 and Fig. 2.7). For printed output, there are line printers, typewriter terminals, and report-writing terminals. Line printers can have very high rates of output, while typewriter and report-writing terminals are usually much slower. Graphical output is possible on many devices, but the most common types are video terminals, storage scope terminals, and x-y plotters. Video terminals (or cathode-ray tubes) require either a microprocessor or CPU attention for ‘constantly refreshing the screen in graphical mode, but allow real-time motion of the graphics and even multicolored graphs. Storage scope terminals require no CPU-screen refreshing, but require that all or most of the screen be erased before a new graph can be drawn, Both types of graphics devices Figure 2.7 Some typical communication peripherals A: Lineprinter,B: Repor-writng terminal, C: ‘Video Terminal, D: Storage seope graphics terminal.22 ADVANCED PROCESS CONTROL, ccan be linked to hard copy units to provide a paper copy of the screen image. Plotters are of many different types, but most provide report-quality finished graphs. Input communications devices include ppewriter, video, and storage scope terminals as well as punched cards, paper tape, and magnetic tape. Removable disk cartridges, which were discussed above, may also be considered to be 1/O communications devices. For long-range 1/0, it is possible to com- municate with the computer over telephone lines, usually through a terminal at a remote location. ‘The rates of communication of these various devices are shown in Table 24, Modern minicomputers are bypassing the requirements for punched cards (and keypunching) by having large numbers of low-cost terminals on-line in a timeshared mode. In this way, all program input can be made directly into the computer with very small load on the CPU. Similarly, paper tape is losing ground to magnetic tape and disks as an input/output and storage medium, largely due to its very slow transfer rates and mechanical problems. 2.3 DATA ACQUISITION AND CONTROL, In order to use a real-time computer to implement a process control algorithm, ‘one must be concerned with the hardware realization of data acquisition and control. There are many different computer configurations which will accom- plish these goals. One may use a large central real-time computer linked to many labs or processes, or have a digital computer network involving many smaller microcomputers, each dedicated to a specific process or lab. We shall begin this section with a discussion of such computer structures. Computer Data Acquisition and Control Networks Depending on the specific application, a wide variety of tasks may be required from a data acquisition and control network: control of plant processes or pilot plants, servicing of terminals in various labs and control rooms, logging of data from analytical and/or research laboratories, interfacing with analog computers for hybrid computation, etc. To illustrate some of these applications, let us consider the network used within the department of chemical engineering at the University of Wisconsin and depicted in Fig. 2.8. The hub of the network is the PDP 11/35 minicomputer described earlier, functioning under a multiuser, time-sharing operating system. The links to the various remote stations are basically of two types, either with or without a remote microcomputer. At resent the Méssbauer spectroscopy lab, the bioengineering lab, the three pilot plant locations, and the remote terminal links all come directly to the central ‘minicomputer without involvement of a microcomputer. The polymer physics lab and the Raman spectroscopy lab, on the other hand, have PDP 11/LSIINTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL 23, Sitersnd — Fy —_— mcronory ‘eral ome toon ‘chine tb eee a U Lif jal Sere ny eemcome) Ehoritny Lam Aston L__} sensing oto ete Biocon Figure 28 The Computer Data Acquistion and Control Network atthe University of Wisconsin, Table 2.6 Capabilities of the computers in the computer network shown in Fig. 2.8 Disk Magnetic ape Memory storage storage ‘Approximate Computer (words) (words) (words/tape) Computing speed cost PDP 11/55 128K —25million 1Smillion Floating point: $90,000 system add 1 shown ia ive, 4s Fig. 22 PDP U/Lsl 32K 256K - Floating point: $ 9,000 system add, 60s; shown divide, 160 ps inFig.2924, ADVANCED PROCESS CONTROL Figure 29 LSi 11 microcom- puter with video terminal, twin floppy disks, 32. K words memory, A/D end D/A con verter, and real time clock microcomputers, which provide local control of the instruments and acquire and store the raw data. Through the links to the central minicomputer, these microcomputers are able to call upon the large file storage (disks), permanent data storage (magnetic tape), and computing power of the larger machine for data manipulation. Table 2.6 illustrates the capabilities of the various computers in the network, while Fig. 2.9 shows one of these PDP 11/LSI microcomputers.* Note that while the microcomputer is relatively inexpensive and has a moderate amount of memory and mass storage, computing speed for the LST 11 is 40 to 60 times slower than that for the 11/55. Thus itis better to use the 11/35 for any substantial numerical calculation. Through different, but compatible, oper- ating systems, the computers in the network are able to transfer programs, data, calculated results, etc, back and forth so as to optimize the performance of the entire computer system. ‘Through a link between the 11/55 and the analog computer system, hybrid ‘computation may be carried out. As noted in Fig, 2.10, this requires A/D and D/A conversion to handle the analog computer inputs and outputs. Some scaling may also be involved because the voltage ranges of the two computers may be different. For example, the University of Wisconsin network has an analog computer operating over the range 100 V to +100 V, while the A/D and D/A converters have the range ~10 V to +10 V. This means that the analog input signals must be amplified by a factor of 10 and the analog output signals divided by the same factor. The most common application of this hybrid computation facility is for control algorithm testing. A process (distillation column, chemical reactor, etc.) may be readily simulated on the analog com- puter and will respond in real time in much the same way as the actual process. This allows one to program the computer control algorithm on the digital + Microcomputers similar to the LSI_ 11 are often used as indutrial process control computers ‘where relatively small numbers of control loops ae involvedINTRODUCTION TO ONLINE DATA ACQUISITION AND COMPUTER CONTROL 25 PDP 11/85 dtl computer ° KE rey Veraviov “oveveriov op outpue analog inp (32 amptten) Figure 2.10 Analog-igital links for bybrid computation, ‘computer and test the performance in real time with the help of the analog computer, Examples of this approach will be given in forthcoming chapters ‘Transducers for Data Acquisition and Control Signal Output It is useful to consider some of the transducers more frequently encountered in the process industries. A representative sample of data acquisition sensors for some of the more common process variables is listed in Table 2.7. Depending on the sensor chosen, the primary output can be either digital, pneumatic (air Pressure), current, or voltage. Because the digital computer can receive only voltage or digital signals, the pneumatic and current signals must be converted to one of these acceptable inputs. In practice this involves special transducers for pressure/voltage conversion (P/E transducer) and current/voltage conver- sion (1/E transducer, which is usually just a high-precision resistor). In a similar way control system actuators come in several different forms. Table 2.8 includes some of the most common types. Notice that in the process industries, the most frequently seen actuators are for flow control. This is because heating and cooling processes often involve such factors as coolant flow-rate adjustment, steam injection rates, or furnace fuel flow rates. Similarly, Pressure adjustments usually require controlling inflow or outflow from a vessel, 88 does liquid level control. Although the digital computer outputs are limited to digital or voltage signals, some of the control actuators require current ot Pneumatic inputs. Thus special transducers are needed to provide for volt- age /pressure (E/P) and voltage/current (E/1) conversion.Table 2.7 Some common data acquisition sensors Process variable Device Oviput 1. Temperature Thermocouple mV Cheap, ragged Resistance sensors (es, thermistor) Oxcilating quartz crystal Radiation Pyrometer 2, Pressure Diaphragm Bourdon tube 3, Flow Pitot tube 4, Liguid level oat Nuclear radiation Photoelectric Head pressure Comments Resistivity High sensitivity Ovcitating High sensitivity frequency nv For high-tempera- ture applications Capacitance Low presure Inductance Both low and ih pressure Photoelectric Resistance Reluctance Stain gauge —_From'ow to very high Prenoeectic pressures Eat Uliesiow pressure Dittereni pressure Dittereniat presure Mechanica 7 defection Rotational speed Generated emt For difficult flow measurements Nuclear For difficult flow radiation ‘measurements Doppler shit Resistivity High precision Mechanical detection Resonant trequeney Resistance Capacitance Resistivity Radiation imtensity Voltage Dittereniat pressure‘Table 2.7 Some common data acquisition senors (Continued) Keeper cenors (Continned) Process variable ‘5. Composition Device Poteationetry Moisture content: Hygrometry Payehrometry ‘atomic-emission, seray, electron, fon, magnetic Polarography Conductimetry ‘Mass spectrometry Differential ‘thermal analysis Thermogravimetric analysis Output Comments av Resistivity ‘Weta dey bulb tempera. av Long analysis av Sound velocity Ttesity Many ofthese techniques are not yet suitable for ansine pplications av Resistance nV Heat wans- ferred Mas changes oT * There are so many composition detection devices that only a few are noted here. Table 2.8 Some typical control system actuations Contra action| letrical heating Flow adjustment On-ot signals Device Blectie Furnace Pneumatic valve Solenoid vaive Motordriven valve Variable-speed pump Fuidie control valve Lights, bells Relays, switches Tapa Voltage or current Air pressure Voltage, current, or digital pulse Digital pulses Voltage or current Pressure Digital pulse Digital pulse28 ADVANCED PROCESS CONTROL, ‘Signal Conditioning ‘An additional consideration in data acquisition and control is conditioning the sensor or actuator signal so that it will have high accuracy, have low noise levels, and makes efficient use of the interface equipment. Usually the steps involved in Signal conditioning are multiplexing and amplification of the signals, suppres- sion of noise on the signal, and selection of a transmitting mode for the signal We shall discuss each of these considerations in turn, In some applications, there will be many similar measurements at a remote location, and it is usually most efficient ro multiplex these at the source and ‘transmit them sequentially over only a few lines, As an example, consider 10 thermocouple measurements from a process, each with a voltage signal of ~ 10 mY. First of all, one would not wish to run 10 lines and use 10 channels of our A/D converters just for these signals, so multiplexing is necessary. In addition, such low-voltage signals will be heavily corrupted by noise in transmission and Will have very low resolution when they reach the A/D converter; thus amplifi- cation as well as multiplexing is needed. Figure 2.11 illustrates two possible approaches to solving this problem. Scheme (a) involves amplification of each low-level signal (requiring N high-gain amplifiers) up to ~ 5 V, then multiplex. ing with solid-state relays. By multiplexing high-level signals, one minimizes the effects of noise introduced by the relays; however, this is at the cost of an expensive amplifier for each signal. Scheme (6), on the other hand, multiplexes first and then requires only a single high-gain amplifier. This is more efficient, but requires care in selecting a low-noise-level multiplexer. Such a device usually equires relays with mercury-wetted contacts in order to avoid corrupting the low-level voltage signal. Both types of multiplexing schemes are known to be successful, and one of them will be illustrated by an example given below. Noise suppression, either before or after signal transmission to the computer, is also an important part of signal conditioning. There are times when relatively high-frequency noise, such as 60 Hz coming from power lines or motors, appears on the signal. Fortunately, in the process industries, such high frequencies are yery rarely contained in the desired signal; thus filtering to remove these frequencies can be a very successful means of noise suppression. A simple Jow-pass filter with frequency response characteristics like those shown in Fig. 2.124 will often suffice. Other common types of filters include high-pass filters, which reject low frequencies, and notch filters, which pass only signals within 4 specified frequency band. Figure 2.13 shows an example of the performance of a low-pass filter applied to a thermocouple signal afflicted with motor noise. The last part of signal conditioning involves a decision on how the signal is to be transmitted to the computer. There are several alternatives available for signal transmission: 1. Voltage signals—These work well over short distance cables (up to ~ 300 ft) ‘where voltage losses and cable capacitance do not cause signal deterioration.INTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL 29 Senor sale Tomi 500 Xamp, 1o—— Seaueniy sutiperer vetage (o) Amplitcaton followed by mulpesing Senor sits Slomv “Awpiiiog 00x Smale Matiponer (6) Malling foowed by aplication Figure 2.11 Two different strategies for multiplexing and amplifying low voltage signals. 2. Current signals—This method works well with hardwire connections over longer distances but requires special transmitters and I/E conversion at the receiving end, 3. Digital signals—These signals can be transmitted most easily and are the natural choice when the original signal is digital, Even voltage signals may be transmitted in a digital mode through D/A conversion at the sending location and A/D conversion at the receiving end. Digital signals have the advantage that they are less susceptible to noise problems and can be transmitted over extremely long distances through telephone connections. OF course, the selection of transmission mode should be made based on the Particular application, noise levels to be encountered, and distances to be covered.[390 ADVANCED PROCESS CONTROL Fregueney (a) Low pt fier ‘Amplinde 7 Frequency (0) Hie ae fter Amplitude Frequeney Figme 212 Frequeney response char- (0) Noch er acteristics of some common filters. r 1 Te Se Low pase To ‘Themacoope oe ae caper sien! Son Figure 2.13 Performance of « low-pas filer.INTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL 31 Modes of Computer Control ‘Computer control is usually carried out in one of two modes: supercisory control or direct digital contro! (DDC). Supervisory control, illustrated in Fig. 2-14a, involves resetting the set point of the local controller according to some computer algorithm, Thus the computer control scheme need only supervise and coordinate the actions of the local controllers. Direct digital contol, by contrast, requires that all the controller action be carried out by the digital computer, Measurements are sent to the computer and compared with the set point, then the computed control action is transmitted to the actuator. This is illustrated in Fig. 2.146. For DDC the computer samples the flow measurement at discrete instants of time and sends as control signals step changes in the control valve stem position. The time interval Ar between samples and controller changes rust be chosen with care. If Ar is taken too large, the controller performance will deteriorate, while having Ar too small will put an unnecessarily high [4 9 Seri om 2 1 AH] re Yaseen Zz (a) Superzry conte irre 2 [St ir. (6) Dic git onto Figure 2.14 Supervisory and direct digital control of a uid flowrate32 ADVANCED PROCESS CONTROL Computational load on the computer. Both supervisory control and DDC are in Wide use in industrial applications, and both may be readily used to implement modern control strategies. In most processes, the local contollor dynamics (whether analog or digital) are short compared with the time constants of he Trented ia a eam; thus most sophisticated control strategies are imple, roaat gi, 8 cascade fashion—involving programmed set-point changes infec seat controllers. Hence, it matters lie whether these local contoecs <°o rales or digital. As an example, suppose the flow-rate control loop in Fig. 214 algeria tem into heat exchanger. Suppose further that an optimeadios algorithm calculates the optimal steam flow-rate program FO). Clearly this Noamralve fee implemented by either supervisory control or DDC when the Be have provided here only a brief summary of the considerations important ia Ghline data acquisition and computer control. For more detail, the reader Should consult the references at the end of the chapter 2.4 SOME EXAMPLES T,llusrate some of the principles discussed in the earlier sections, several Sragible case studies shall be presented. It is hoped that these will pive the ‘eadet a practical grasp of what is involved in interfacing processes to reali computers, Computer Control of the Pressure in a Gas Storage Tank sire digital computer must be programmed to receive the pressure measure- cramped calculate control signals based on some controller algorithm: Fy ‘example, a simple PID controller in discrete form would be Ho) = 06+ & {Lu ra)] +23 bv - 200] + 210 - wal} eanINTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL 33 one 3-15 pa Sma cov | Boe} = PO" = osomv orev reo be ae cuter: one Figure 2.15 Interfacing forthe gas storage vessel. where u(i,) is the control signal at time %, uy is the zero-error control signal, K. is the proportional controller gain, 1, is the integration time, and rp is the derivative time. The measured output (pressure in this case) at time f, is denoted by (4), and y4 is the controller set point. The quantity Ar is the sampling/ control time interval, and is usually selected to be short compared with the time constants of the process. « soz Tank presse Tank reste (pi) ‘atone ° 150 300 Tine (6 Figure 2.16 Dynamic behavior of the gas storage vesel under DDC control.34 ADVANCED PROCESS CONTROL The control signal, which is transmitted as an analog voltage, must be converted to a 3- to 1S-psig air-pressure signal in order to drive the pneumatic valve. This could be done in one step with an E/P transducer; however, to make use of existing equipment, this conversion is done in two steps, as shown in Fig. 215, Alter installation, the transducers must be calibrated. For example, 0 to 60 psig corresponds to 0 to 10 V analog voltage for our system, while 0'to 10 V analog control signal corresponds to 3 (full shut) to 15 psig (full open) for the control valve. Because the analog signals were sent over shielded cables for relatively short distances and were not in the vieinity of power lines or motors, ‘no noise suppression devices were needed in this application. Figure 2.16 illustrates the dynamic behavior of the implemented control scheme, showing both the measured tank pressure and the computer-specified control valve Position. The control time interval At was taken to be only a few seconds, so that the DDC step changes in valve position appear as a continuous curve in the plot. The actual valve stem position does not exactly follow the specified control signal due to the dynamics of the valve itself. Nevertheless, the control system performs adequately, as might be expected. Computer Control of an Instrumented Steel Ingot A more complicated scheme of process interfacing can be seen in the following example, It is desired to control the temperature distribution of a cylindrical steel ingot by manipulating the power input to each zone of a three-zone furnace. This isa laboratory model of a metallurgical heating furnace; itis much ‘more heavily instrumented with thermocouples than is possible in practice. This instrumentation is to allow an absolute measure of the temperature distribution. in order to evaluate the performance of proposed control algorithms. A photo. sraph of the furnace system is shown in Fig, 2.17. Thirty-two thermocouples are Placed axially and radially in the ingot, and each has a 5- to 10-mV voltage aby 217 ‘An instrumented ste! ingot in a thre-zone furnace: A, furnace; B, thermocouples; C, furnace heater relays; D, multiplexer and heater controller; E, link t computertet P2H Peon m0} mons jus onus pee wonenboe war andi36 ADVANCED PROCESS CONTROL signal, These signals must be multiplexed and amplified before they are sent to the computer, Similarly, the control signal for each zone of the three-zone furnace must be converted to a power input for the electrical heaters. Figure 2.18 shows the data acquisition and control signal design used for the system. The 32 thermocouples are passed through a reference temperature bath® land then to an interface device. The interface device multiplexes the 32 signals through 32 mercury-wetted relays, which introduce very litle noise to the signal. ‘The multiplexing is carried out through digital logic and under digital signal control from the computer. Thus the multiplexing rate and sampling interval are under program control. After multiplexing, the signal passes through a high- quality amplifier with a gain of 500. This brings the 5- to 10-mV signal up to 2.5, to 5 V, and this may be transmitted and processed by the A/D converter without undue error. Through multiplexing, the 32 thermocouple signals require only one channel of the A/D converter—a very efficient use of computing resources. To allow each relay of the multiplexer to properly settle before reading the signal, one must not switch the multiplexer too fast. In the present ‘application it was necessary to allow 1.5 ms per relay for reliable multiplexing. Nevertheless, this allowed all 32 thermocouples to be scanned in less than @ tenth of a second. For most process control problems, where the process time constants are minutes to hours, such multiplexing times are negligible. ‘The control signals for the three-zone furnace are sent through three channels of the D/A converter as analog voltages in the range 0 to 10 V. The interface device has digital logic, which converts the input voltage to an arcless relay controller which controls the power input to the heaters. The conversion is, linear, with 0 V control signal corresponding to zero power input and 10 V control signal causing full heater power input. The interfacing scheme shown in Fig. 2.18 performed well and allowed a detailed study of state estimation and control algorithms for the system. One of, the case histories discussed in Chap. 6 demonstrates the performance of this interface design. REFERENCES 1. Andrew, W. G.: Applied Isirumentation i the Process Indusrier, vols 1, 1H, Gulf Publishing, Houston, 174. 2. Considine, D. M.: Process Inuuments and Controle Handbook, 2 ed, MeGraw-Hil, New York, 4, 3. Ewing, G. W.: Instrumental Methods of Chemical Analysis, MeGiraw-Hil, New York, 1975 4. Finkel, Jules: Computer-Aided Experimemation, Wiley, New York, 1975 ‘As an alterative to ice baths, there are inexpensive electronic compensators which work ually welINTRODUCTION TO ON-LINE DATA ACQUISITION AND COMPUTER CONTROL. 37 5. Holman, J.P: Experimental Methods for Engineers, McGraw-Hill, New York, 197. 6 Korn, G. A: Miniconputers for Engineers and Sciensts, McGraw-Hill, New York, 1973. 7. Norton, H. P. Handbook of Transducers for Electronic Measuring Systems, Prentice-Hall, Englewood Cis, NJ, 1968, 8 Perone, . P, and D. 0. Jones: Digital Compuiers in Scienifc Instrumentation, McGraw-Hill, New York, 1973, 9. Theis, D. J: Datamation, p.73, (Feb. 1977), 10, DataPro Reports on Minicomputers, DataPro Research Corp, October 1978 LU, Harrison, TJ: Handbook on Industrial Computer Control, Wiley, New York, 1972 12. Harrison, TJ: Mnicomputers in Inducril Control ISA Pub, Piteburgh, 1978, PROBLEMS 21 Convert he following binary numbers o octal numbers (@ 100 111 000 111 101 (@) 101 010 101 110 oF (© M1 061 110 100 ont 22 Convert the following octal numbers to decimal numbers: (@) 747 (©) 100 (40 (a) 556 123 A Fortran program consists of 100 lines with approximately 28 characters per line, Estimate the ‘number of words of memory in a 16-bit machine required to store thi Fortran code. 24 In responding to interrupts from various peripheral devices, « computer manufacturer recom- sends the following priority levels: Terminal—4 Line printe—4 Realtime clock—7 A/D conver Disk drive—7 Magnetic tape unit—7 Diseuss why some of these peripherals are assigned higher priorities than others. 25 A computer has a 12-bit A/D converter with —10-V 10 +I0V span. Determine the analog ‘input voltages and relative error forthe following integers coming from the converter: (@) 570 (@) -960 (@) 20d) 25 26 Prove the answers to Prob. 2 for a 10-bit A/D converter with a the following intege (@) 450 (6) -20 (@) 10 (dy 25, S.V to +5.V span and for 27 Ic is desired wo output the following voltages through a 12-bit D/A converter with a —10-V to ‘+10. span. Determine the integer that should be loaded into the D/A in each case. @-1v (@ 427V ( +87V @ -6v 28 The folowing analog circuit has been suggested as a simple low-pass Miter. Wete down the ferential equations corresponding to this cireuit and indicate how this circuit might accomplish the filtering action. Can you suggest values for R and C that will eliminate 60-Hz noise without sisturbing the I-He signal?[38 ADVANCED PROCESS CONTROL Input Outpt 129 Two computers ar linked over a 300-baud telephone line. One computer would like to transfer Fortran program to the other computer over ths telephone link. The Fortran program in question has about $00 lines of code with approximately 30 characters per line. Estimate how long it wil ke to transfer this program, An engineer sugges connecting the two computes via & 9600-baud cable link. How long would the program transfer take inthis case?(CHAPTER, THREE CONTROL OF LUMPED PARAMETER SYSTEMS 3.1 INTRODUCTION In this chapter we shall consider the control of lumped parameter processes, processes described by ordinary differential equations. These include both single-variable and multivariable dynamic systems. We shall begin with a discussion of some of the key concepts and practical difficulties encountered in lumped parameter control system design. Some of the control system design procedures available for linear multivariable systems shall be presented and illustrated with examples. Following this we shall introduce optimal control theory, which represents one approach to control system design, and which provides a theoretical foundation helpful in understanding material to be ‘covered later in the book. Finally, we shall discuss special techniques which may be used for nonlinear dynamic systems. To provide a very simple example of the type of problem we shall be attacking and to clarify a number of concepts, let us consider the simple, well-stirred, steam-heated mixing tank shown in Fig. 3.1. It is used to mix and preheat a recipe of reactants before passing them to a chemical reactor. Now let us suppose that we wish to design a scheme for controlling the effluent temperature T(J) to a desired value T, by manipulating the steam rate Q(¢) with valve v. ‘An open-loop control scheme for such a process would involve programming the steam valve position over time without benefit of feedback information, such as a tank effluent temperature measurement. Such a scheme might be useful if @ very good mathematical model of the process were available, €.2., a FG-7), OW a ¥— * Voc, 70) = To G12) and one had the start-up problem of driving 7(1) from the initial condition Ty to GD) 940 ADVANCED PROCESS CONTROL stem Openoor A} —FA, set ew A se + Figure 3:1. seam-heated mixing tank. Ge ni yl Tn opp ec pay ol 5 ere tee Fant cbc Pope on gel be eae ain a 2 eee Se rom de oc wan el ee eo nen eee foe cae eee ee ay one peice See ee a ee aay tie te Solgar og a ene epcieetleaeat triers (PID) controller given by a= aealen- met fin na +X Danny wes tg ty he ont ance Hove vl lio Tene cn une chars oer be al Le ee ae eh eee eter ol schemes for processes described by ordinary differential equations. 3.2 LINEAR MULTIVARIABLE CONTROL SYSTEMS Perhaps the most commonly encountered control system design problem in the process industries is the design of multivariable control systems. If the process under study is linear, a very general model in the time domain (so-called state‘CONTROL OF LUMPED PARAMETER SYSTEMS 41 cariable notation) is Ba acs Buta X(t) = Xo (32) y=Cx (322) where is a k vector of disturbances, is an n vector of states, is an m vector of controls (manipulated variables), and va v2 yo w is an [ vector of outputs (those states or combination of states which can be measured). The matrices a a a bu ba bm] an a by 2 . 2m a-|® fy male ” Gn 4, Ont Pam cu te ein Mm Me Ye aed Sol eel ™ “ns ve42 ADVANCED PROCESS CONTROL can be either constant or time-varying. This very general model shall be used as a basis for much of the discussion which follows. Tris important to note that Eqs. (32.1) and (322) have analytical solutions which are useful to know Inthe case ofan autonomous sistem (ie, where A, B, Gand T are constant matrices, one may take the Laplace transform of a, G21) to viele SIR(s) — x = AX(s) + Bu(s) + Pd(s) ¥(s) = CR(s) Solving for X(), we obtain 265) = (1 ~ A)"[xp + BAK) + FH) ] 623) Now, using the convolution theorem of Laplace transforms to invert this expression, the analytical solution takes the form x(t) = MWg + ot) + Ta(r)] dr (324) where the exponential matrix e** must be evaluated. Let us note that X = e** is the solution fo the simple homogeneous matrix differential equation ax Fat ax x0) 25) where X is ann Xm matrix. If the eigenvalues of A are distinct, then there is a canonical transformation [1] A= MAM"! (3.26) relating the matrix A to its diagonal matrix of eigenvalues » o oy G27) ° A and matrix of eigenvectors M. Substituting Eq. (3.2.6) into Eq, (3.25) yields AO) AQ) MIXO=M-! G28) which has the solution M"!X = eM X= ec! = Mew! (3.29)‘CONTROL OF LUMPED PARAMETER SySTIMS 43 where e* is the diagonal m x m matrix 3.2110) ‘Thus one direct way to evaluate e is through Eq. (3.2.9). In the event that the matrices A, B, C, and I’ are time-varying, the system represented by Eqs. (3.2.1) and (3.2.2) is nonautonomous, and an alternative solution must be used. In this case, the analytical solution takes the form [2} x(t) = (6, f0)X +01, fm te) *[B(r)u(r) + L(A(r)] dr (3.2.11), where (r, 4) is an n Xn time-varying matrix known as the fundamental matrix solution. The fundamental matrix solution arises from the solution of the equation dos, 1) = a(1) (6,19) Mt ta) (3.2.12) Example 3.2.1 To illustrate how these general results apply to a particular ‘multivariable system, let us consider the stirred mixing tank in Fig. 3.2 with cross-sectional area A,. There are three streams entering the tank: (1) a hot oe oe | rg | fo ay Daturbance stream ToFe Figure 3.2 Stirred mixing tank requiring evel and temperature control.44 ADVANCED PROCESS CONTROL stream at temperature Ty with adjustable flow Fy, (2) a cold stream at temperature Te with adjustable flow Fc, and (3) a disturbance stream from ‘another process unit with variable temperature T, and flow F,. The tank is ‘well stirred, with liquid outflow determined by the liquid height in the tank, h, ie, FH) = Kh'/?, The modeling equations arise from material and energy balances: aft Ge = Fx t Fo t F~ Ui) AA ODD m pel Fyn + FeTe + Fala FUT] 02.13) Clearly the tank mode is nonlinear; however, itis possible to linearize the Stontins about the desired operating point hy, T, by wsing Taylor series apansions rancaled afer the first-order terms! 1k FU) = FU) +5 Gia + FUT = FUA)T, + FAM(T ~ 741 AT ah, + A(T - 7) + Th = h) +o “214 Fil ™ Fan + Fa(Ta~ Ta) + TalFa~ Fa °° > If we then define Fy Fe Fay Tu 5 the steady-state values of Fy, Fes Fay T, corresponding toh = hy, T= 7, that is, satisfying 0 = Fay + Fes + Far ~ FUH) O= FT + FesTe + FaTa — FT, (3.2.15) then one can use the deviation variables mah-h me TOT mm Fy Fin we Fe-Fa =F Fe = T-T (216) to obtain the linearized modeling equations 2.17)(CONTROL OF LUMPED PARAMETER SYSTEMS 45 Hitec Sa ty = Tm + (Te — Tita + (Ta = Tt + Fas — Flhs] (218) Thus, if we define the vectors (3219) then the model for this system is in the form of Eq, (3.2.1). If both the state variables are measured, then Hy aps Lo rfi]-[R] mt ero a] To use the analytical solution, Eq, (3.2.4, for this example problem, let us note that the mats A is diagonal and that the characteristic equation |A-aI] <0 Yields the eigenvalues 1A) ah = (3.2.20) Hence A is already in canonical form and o([e 2] 6221)46. ADVANCED PROCESS CONTROL, Let us now evaluate the response of the system to a unit step change in 1, and w, at ¢ = 0, that is u, = 1, u, = 1. Further, let us assume xOmm=[Te] and de dno In this case Eq. (3.2.4) becomes x(1) = eMxy + f “eM Bu(s) ds (3.2.22) or ay | Ot fee gt Ay fant eI + f'e™ ( Je Evaluation of the integrals yields the analytical solution: x)= oof Reo rab! . | - z ae soon pao + AA FES 2 (1 oxp| |} (223) ‘Time-Domain versus Transform-Domain Representation In considering the dynamics of multivariable linear systems with constant coefficients, one may choose to do the analysis either in the time domain, with models of the form Ba ac+ But Ta (fo) = Xo (8.2.24) yacx (3225) or in the Laplace transform domain, involving transfer functions in the form ¥(s) = G(s)i(s) + G(s)a(s) (3.2.26) Here the overbar denotes the Laplace-transformed variable and the matrices G(s), GAs) are the multivariable transfer functions relating the system control‘CONTROL OF LUMPED PARAMETER SvsTEMS 47 variables ii(s) and disturbances d(s) to the system output, i¢., Buls) Bilt) 7+ Bil) 66) =] 80) HO a) Bils) Bin(5) (a2 gis) gis) -+* site(s) G4) -| BO 8) Hho Bits) Bals) Note that the state variables x are not used in the Laplace-domain representa- tion, ‘Some control system design procedures are easiest in the first representation, while other algorithms are more convenient using the second. Thus the control system designer should have the capability to quickly switch back and forth between the two formulations. Let us discuss how this may be easily done. ‘The transformation from the state space (Eqs. (3.2.1) and (3.2.2)] to the transform space is unique and simply done by taking the Laplace transform of Eqs. (3.2.1) and (3.2.2) with xp = 0, to yield ¥ = C(sl — A)"'Bu(s) + (st — A) 'Ta(s) (3.2.28) Hence G(s) = C(t — A) 'B (3.2.29) G(s) = CGI = AT (3.2.30) ‘There exist standard computer programs (see Appendix A) which can efficiently perform the operations in Eq. (3.2.28). ‘The reverse transformation from the transform domain to the time domain is more difficult for several reasons: 1. Because there is more information contained in the time-domain representa tion than in the simple input/output transfer function, the transformation from the transform domain is not unique. There are many sets of Eqs. (3.21), (3.22) equivalent to Eq. (3.2.28). 2. Among the many possible transformations, one wishes to choose one which provides the minimal realization of the transfer function. This means the smallest set of state variables necessary to match the transfer-funetion rela- tionship. Even so, there are many possible choices of a minimal realization.48. ADVANCED PROCESS CONTROL, ‘Among the several different algorithms for finding the minimal realization of a transfer function (see [3-5] for a more detailed discussion) is a most useful technique which works when all the poles of the transfer function matrix are real and distinet. The procedure may be outlined in the following wa Let us conser the transfer function matrix G(s) which can be expanded in ‘partial fraction expansion of the form oo - 3 623) where the -A, i= 1,2,...p, are the poles and the / x m matrix M, is the matrix of residues for the pole —A,, defined by M, = lim, [6 + )6)] (3232) If my, is the rank of the ith matrix of residues, then the minimal state space realization must consist of n state variables, where (3.2.33) ‘The construction of the matrices A, B, and C can be done in many different ways, and, as noted above, the selection is not unique. One procedure is to make ‘A diagonal and of the form A= : (3234) where the diagonal submatrices are given by A, = —M, and I, is an identity matrix of dimension n,. To determine the elements of B and C,note that with A in diagonal form, the quantity (s1 — A)~'is also diagonal, so that Eq. (3.2.29) may be rewritten in the form of Eq. (3.231), where M, will be‘CONTROL OF LUMPED PARAMETER SYSTEMS 49 composed only of the elements of C and B: M,= Deby f= 1,2, (3.2.35) where ¢, is an / vector and by is an m vector. Then the vectors ¢,, by are one possible choice for the columns of € and BY, respectively, and these would take the form Oe (3.236) where, of course, B is ann X m matrix and Cis an / X m matrix. If in addition we must invert the disturbance transfer Function G, = C(st— A) (8.230) we must now choose the m x k matrix I° to obtain the proper transformation, ‘This can be done by expanding G, in the form of Eq, (3.2.29), where the poles of G and G, must be the same:* 2 Me a-2 str 237) Here the matrix of residues Mf may be expanded as Mix Dewy (3238) * This is not really restriction because a common denominator may be defined which contains all the poles of G and G.‘50 ADVANCED PROCESS CONTROL where the elements c, have already been determined and the k vector y must be selected to form the matrix wh va (8.239) Yes, Let us illustrate these procedures by examples. Example 322 Let us consider the dynamic modeling of a liquid-phase chemical reaction for reacting species A to products. Assume that the rate of reaction is independent of temperature and that the reactor chosen is an adiabatic stirred tank. Composition and energy balances yield the modeling equations dey 1 on gleu- ea) ~ ke FUG 1+ Hey Here we assume 7; is the manipulated variable and T the output, so that by letting x; = cq. X2= Ty ty = Cap ty = Ty and y = T, one obtains « Beato ya cx (3.2.40) ° 5 1{ Be c=[0 1] a D e o‘CONTROL OF LUMPED PARAMETER SySTEMS 51 ‘Then, by application of Eq. (3.2.28), one obtains Gls) = C61 = A)" where s+ o (i= : need and 1 a= 1 ere | GI A" = G70 HGF 17) o selon Thus , : tte CO" Baie Hee] @ = —e— | O40 is the unique transfer function required. Example 3.23 Let us now determine a minimal state space realization for the transfer function, Eq. (3.2.41), derived in the previous example. Clearly the poles are ~(1/@ + k) and 1/0. Thus a partial fraction expansion of, the form of Eq. (3.2.31) gives on setreel-$ +13 3] Since the rank of each of the M, /= 1,2, is 1, a minimal realization requires two state variables. Let us now determine a set of A, B, C for this realization using the algorithm suggested above. In diagonal form, (3.2.42)Zo 6 Fai) ety 8.243) a a is an acceptable choice. Note that this leads to the state space model Clearly y, my, and u, have the same physical meaning as in the previous example, but x, and x, here are not the same, In fact, x, = Jey + T, xy = — Je, here. ‘This example clearly illustrates the nonuniqueness of the minimal state space realization, since the transfer function of Eq. (3.2.41) has been shown to be equivalent to two different state space models, Eqs. (3.2.40) and (3.2.42), 3.2.43). More complicated transformations will be illustrated in the next example problem. Example 3.24 In this example we illustrate the state space representation of ‘4 more complicated transfer function relationship, ¥ = Gils) + G,d(s) (3.2.44) where G(s) = GAs) =| _05 1.0 Making use of Eq, (3.2.31), one obtains Tyo on] +54 2 1 joe: sf) sal 0 where A, = — §,A, = ~ } are the poles. Evaluating the ranks of the residue matrices, one obtains n, = 2, 2; = I, s0 that the minimal realization must be(CONTROL OF LUMPED PARAMETER SYSTEMS 53 of order 3. By using the diagonal form, 0 oO 4 0 A=| 0 - oo and expanding Mi, Ma nth form of Eq, (3.238), then ' 0 m=[1] [oom 0} +(2] (0 004) M, -[!] [0.25 0] ‘Thus B and C can take the form 0078 0 5 a-[°O cou] c=[} °°] os °9 ‘Now expanding G(s), one obtains sun 7[0 ot] * [oas 6] S [odes 3] and expanding the matrix of residues Mf according to Eq. (3.2.38), one obtains: [2] t@ o1+[9] 10 omy : Mj= ] [0.0625 0] ‘Thus r= o 0 Oo oll 0.0625 0 ‘Summarizing, the state space model takes the form aa}? 4 3] BMS Ble oom | [ ] x M 025 0 0 was 4] GL a) a | and is an equivalent minimal realization of the transfer function relation- ship, Eq. (3.2.44), + xe x‘54. ADVANCED PROCESS CONTROL 4 >| Disturbance Figure 3.3 Multivariable block diagram. Multivariable Block Diagrams Having developed the facility to readily convert linear constant-coefficient systems between the time domain and the transform domain, itis now useful to consider multivariable feedback control loops. AS a preliminary, let us discuss the meaning of the simple multivariable block diagram shown in Fig. 33. In the transform domain, G(s) and G,(s) are the plant transfer functions relating the control variables and disturbances to the process outputs, H(s) = G(s)H(s) + G,(s)a(s) (6.245) Here G(s) and G(s) ate defined by Eq, (3.2.27), where the elements take the form* yIL(a + ngs) Ta + efs) for lumped parameter systems. The parameters are as) = (3.2.46) K,—transfer- — (— Af)! — transfer-function zeros function gain (-ef)~' — transfer-function poles Itis interesting to realize that these block diagrams may also be represented in the time domain. In this case, G and G, may be thought of as integral operators. For the linear system equation (3.2.1) with initial condition xp = 0, G is defined as ¥(1) = tg then the system is completely controllable. It is also possible to have systems which are only partially controllable, i.e, in which there are some subsets of initial states xg(f4) which cannot reach every other state in finite time. It is possible to define conditions of controllability for specific classes of systems [2]. For example, if the matrices A and B are constant, then it can be shown [2, 7] that the system of Eqs. (3.2.1) and (3.2.2) is completely controllable if and only if the rank of an n X nm “controllability matrix” L, isn, where 1 = [BIABIA’B: ~~» ja"~'B] (62.60) ‘An informal derivation of this result may be seen by considering the analytical solution to Eqs. (3.2.1) and .2.2).* x(t) = eMxy + fra) ds (3.261) * We neglet the disturbances d() because contolabilty i « property of the system itself‘CONTROL OF LUMPED PARAMETER SYSTEMS 57 where the exponential matrix may’be writen [1] eM mTHAL +A + 6262) which when combined with the Hamilton-Cayley theorem* (1] leads to the finite series representation Mm el + Att (AN += tq (AN! 263) Upon substitution of this into Eq, (3.2.61), one obtains x0) = e+ [Lo + ole —9AB + (6.2.64) or ails) rey “|ds (8265) ‘u(s) Now the concept of controllability means that the control w is capable of influencing all the states x through the integral in Eq. (3.2.68). Therefore the system is controllable if and only if the integrand in Eq. (3.2.65) allows the influence of w(#) to reach all the states x(s). This requires the m x nm matrix transformation eee [ BiAB: to have rank n, ‘Note that output controllability conditions can be found by multiplying Eq. (8.2.64) by C to yield WOmCx= Ceding + f"LegcB + a(t S)CAB + - + eat = 5)" CAB Ju(s) ds 3.2.66) Thus, by the same arguments, the controls u(t) must influence all the / outputs ¥() for ouput controllability. This means that the outputs y are completely ‘controllable if and only if the rank of the / x mm controllability matrix L? is J, where L: =[CBICAB: ... |CA"~'B] (3.267) * This theorem states that every matrix satisfies its own characteristic equatios lnfinive series can be reduced to an m-term series,‘58 ADVANCED PROCESS CONTROL Controllability conditions for the case of a linear nonautonomous system of the form of Eq. (3.2.1), when (1), B(#) are known functions of time, specify that the nonsingularity of the m x n matrix M(to 4) = [C6 4) "BOB 0)” ]"" dt (3.2.68) is necessary and sufficient for controllability. Here the m Xn matrix ©, 1) is the fundamental matrix solution defined by (6, f9) = A(NM( fe) WLfp fe) = 1 2.12) ‘The proof is straightforward (2), but shall not be given here. Note that since the integrand of Eq, (3.2.68 is positive semidefinite, itis sufficient for controllability ‘that the integrand be nonsingular for any instant of time fo <1 < tp Let us illustrate these points with a few examples. Example 3.2.5 Let us consider an isothermal continuous-stirred tank reactor CSTR with the irreversible first-order reactions A” B*? C taking place. “The rates of reactions are given by nahin no kee where ky and ky are constants. ‘The modeling equations for this system take the form dea Go = eye) — VUkies)— e(0)= eo VSB = Fle ~ 9) + Vlaea~ byes) 660) = enn tis required to control c,, ¢g as close as possible to a desired set point cy, pq by adjusting cy, and Cy, the feed concentrations of A and B. Let us now define fF v ea Cat where cis an arbitrary reference concentration of A. In this instance, the ‘modeling equations take the form Sy dt x, a ‘Thus one wishes to control the reactor outlet concentration x), x; by = (+ Dada +m a(t) = 0 Dax; — (1+ Dax +a 2) = 0(CONTROL OF LUMPED PARAMETER SYSTEMS 59 adjusting the feed concentrations 1, us. By putting these equations into the general form of Eq. (3.2.1), the controllability criterion may be tested by noting that = (1+ Da) ° 10 - Be ‘ [ Da, ~ (1+ Da) Oo. and the “controllability matrix” is = (1+ Da) ° Da, ~ (1+ Da) Clearly the rank of L., is two for this second-order system, so that the system is completely controllable. Example 3.2.6 Let us now consider the question of controllability of the reactor in Example 3.2.5 with the modification that 1, = ¢y//Cy = 0 and. only 1, the feed concentration of A, may be manipulated to control the reactor. In this case A is given as before, but B and L, take the form veo] Ls ] Da, ‘Again the rank of L, is two, so that the reactor is completely controllable with only control variable u,. However, the controller performance would surely be poorer than for the case when both u, and u, are available. Stabiirabil ‘A much weaker condition than controllability for a system is the property of sabilizabilty (2). Stabilizability is the property that all the unstable modes of the .pstem Eq. (3.2.1), can be made stable by controller action. This means that any positive eigenvalues of A may be made negative by controller action. Clearly then, any system with A having all negative eigenvalues is stabilizable (in fact, ‘ven without controller action). In addition, any system which is controllable is automatically stabilizable. In the case of the constant-gain proportional feedback controller on the state variables u(s) = —Kx(s) (6.209) where K is an m Xn feedback matrix of controller gains, the system of Eq. G21) becomes ax i= (A~ BK) 6.270) ‘The system would then be stabilizable with such a feedback controller if and only if there existed a combination of feedback gains ky which would cause the ‘eal parts of all the eigenvalues of (A — BK) to be negative,(© ADVANCED PROCESS CONTROL With a proportional controller on the output variables a) = —Ky(0) (2.21) the system of Eq, (3.2.1) takes the form (A BKOx G27) In this case, stabilizability requires that the real parts of all the eigenvalues of (A ~ BKO) be negative for some selection of the feedback gains k,,. ‘We shall illustrate these points with an example problem. Example 3.2.7 Let us consider the reactor problem of Example 3.2.5 except now ¢4 is fixed and only cyy may be adjusted to control c, and cy. In this, case let us define EF “7 c, T+ Da 7 kv fy Dye "F “> o*TeDa, so that the modeling equations take the form Bie (1+ Dade, mi) = x0 (273) 2 oda, (1+ Dad ty ll) me 274) In this instance, the A matrix is given as before and 0 | oO »-[1| Le-li) -(+da) Here the controllability matrix L, is singular and the system is not com- pletely controllable. Only the state x, can be controlled by 1, and x; is completely free. However, from the solution of Eq, (3.2.73), one sees that (0 = Xo exp[ ~ (1 + Da)(e~ 4)] so that the eigenvalue associated with x is always negative and x; is stable. ‘Thus the system of Eqs. (3.273) and (3.2.74) is stabilizable. In practical terms, this means that while both concentrations A and B cannot be controlled by adjusting the feed rate of B (only B can be controlled in this manner), the concentration of 4 is stable and will approach the steady state unaffected by whatever control action is taken. Thus one could successfully design a control system for Eqs. (3.2.73) and (3.2.74) even though the system is not controllable. Let us review the implications of controllability and stabilizability in the light of these three examples. First, a control system can always be designed for a completely controllable system and is sometimes impossible to design for one(CONTROL OF LUMPED PARAMETER SYSTEMS 61 which is not completely controllable. If the system is stabilizable (but not Completely controllable) and the uncontrollable eigenvalues of the system are ‘sufficiently large and negative, then an acceptable control system design is possible. However, if the system is nor stabilizable, then control is generally impossible. For Examples 3.2.5 to 3.2.7, the use of control variables cy ¢py together ot cq alone will allow the CSTR to be completely controlled, but cy alone cannot completely control the reactor. Physically this is due to the fact that there is kinetic coupling from c, to cy but not in reverse, ie, the rate of formation of B depends on cy, and thus the control of c, by C4 is sufficient to control cy as ‘well. In contrast, the rate of formation of A is independent of cp, and thus controlling cy exercises no control over c,. Had the first reaction been reversible in Example 32.7, then the reverse coupling would exist and the system would have been controllable with ¢yy alone. Normality A stronger form of controllability is called normality [2]. A system is said to be znormal if each element of the control vector u alone will achieve controllability. ‘This will be true if and only if the normality matrix L. L, =[bjAbja7B!- + ja") (2.75) has rank n for alli, where b,, i = 1, 2,..-,m, are the columns of the matrix B. ‘Thus the CSTR control system discussed in Example 3.2.5 is completely contro- lable but not normal because u, alone will not cause the system to be controlla- ble. For scalar control variables, normality and controllability are identical Properties. The Interaction Problem ‘The fundamental problem in designing multivariable feedback controllers lies in the steady-state and dynamic interactions which occur between the various input, and output variables. If the system had no coupling between variables and the number of control variables equaled the number of outputs to be controlled, then the linear system of Eq. (3.2.1) in the transform domain would have a diagonal open-loop transfer function 8uls) o G(s) = CI ~ A)'B = B28) (3.2.76) o * gis) and for G, in the diagonal form, Eq. (3.2.50), the closed-loop transfer function, Eq. (3.2.58), relating y to y, would then be diagonal and each control loop could be tuned separately by classical methods. Unfortunately, most multivariable systems have significant coupling be- tween outputs and controls, and these pose great difficulties in control system design. To illustrate these problems, consider the following example.(@ ADVANCED PROCESS CONTROL Example 32.8 You are working for XYZ Chemical Co. in their engineering department and the plant supervisor comes to you with a control problem. In the plant is a large binary distillation column, sketched schematically in Fig. 3.4, There are four product streams drawn from this tower, with concentrations of the heavier component in the top three sidestreams denoted by y,, Ya, and yy. The bottoms must satisfy an overall material balance. The compositions of these products at each drawoff point have specifications yyy, Yay and yyy and are controlled by adjusting the drawoff rates, 1, ua, and uy, Currently the tower is run by manual control because the present single-loop PID controllers do not work well. They produce ‘considerable offset and occasionally cause the sidestream compositions to continuously oscillate. Under manual operation there are no oscillations, ‘but there are still significant deviations from the set point which require hours of time and many trial-and-error manual adjustments to eliminate. AS an indication of the problem, the plant supervisor asks you to consider Fig. 3.5, where the operator successfully adjusts y, to a new set point by increasing flow rate 1, and decreasing flow rate u, by an equal amount. This ‘causes the concentration of the less volatile component at sidestream 2, y., a Feet Fe Figure 34 Multisidestream distillation column,(ide pom New set pat Time Figure 35 Open-loop response ofthe distillation column to manual adjustments of wand uy. to decrease to the new set-point value.* However as can be seen, the concentration drawn overhead does not remain constant but decreases also. ‘The concentration y, at sidestream 2 first increases and finally decreases to approximately the old operating point. Thus, by successfully meeting the new specification on yz, the operator has thrown y, off specification and must now work on correcting this by adjusting u,. This begins a whole new round of detrimental interactions, and meanwhile the column is making many tons per hour of off-specification product. Note that because of dynamic interactions which can cause temporary deviations in the wrong direction, after each adjustment, the operator must wait for the transients to die out to determine if this action is a success, The plant engineer asks for your help to develop a better controller for this tower! ‘As a good control engineer, you first request that some modeling work be carried out on the column. Thus step input test data are fit to a rather simple transfer function relating the flow rates u to the sidestream composi- tion (3), ie., (5) = Gils) * The operator uses this approach because he has learned that this usually doesn't cause to be thrown too far off specification ‘Chapter 6 contains & further treatment ofthis example problem,(64 ADVANCED PROCESS CONTROL where G(s) has been found as a 1+ 9s 2 ® 2. os S)*| a ee 27) 2323.2 Ts T+8 T+7s ‘Some interpretation of this model shows that all the responses between flow rates u, and sidestream compositions y, can be considered first-order. However, one should note that in actuality the system is of much higher order (order ~ number of trays between drawoffs). In practice, time delay terms e~*” would often appear, but for smaller columns, these delays may be negligible and have been neglected here. The model shows that there are significant interactions between variables, and in fact, the diagonal terms fare not even dominant. Note that the interactions are one way, i.e, the adjustment of a drawoff rate affects all stream concentrations below it, but none above. This is the easiest type of process interaction to deal with. Let us try to understand the difficulties encountered by the operators. ‘The present column feedback control scheme, which was taken out of service, is of the form shown in Fig. 3.3, where a) 0 0 Gato mm” 0 0 0 gn) represents three single-loop PI controllers. From Eq, (3.2.56), the closed- loop transfer function for the column is 3=9y, (2.78) where T=(1+6G)"'GG,=1-(+66)"' 3.258) and 0785, T+9s o 2081, 0482, ce, =| eu. z 3.2.79) T+8 T+6s ba 238, 238m, 218, T¥l0s T+8 T+7s(CONTROL OF LUMPED PARAMETER SYSTEMS 65 078n, 2.0g1,(1 + 9s)(1 + 68) (+ (1 + 9s + O78, (I + 65 + 0.4822) [ 238n, (jie :) 4.681.833 Ws oe (1 + 9s)(1 + 63)(1 + 75) (1+ 95+ OTe, (1 + 65 + OB (I + 75 + 2.1835.) T¥6s T+ 10s: 4g, T+ 65+ 0483, 2.3g2,(1 + 6s)(1 +78) 2lgy, (1+ 8(1 + 65 + 048m (I + Ts + 21gy) T+ I +2185) (3.280) Observe that there are significant off-diagonal terms in T corresponding to strong dynamic and steady-state interactions in the system. This easily explains how operating difficulties with the present control scheme could arise if the controllers were not carefully tuned. Note that because the interactions are only one way, the controllers could, in principle, be tuned fone at a time, starting at the top of the column. Nevertheless, the control system performance could be greatly improved by a better control system design which would minimize the effects of interaction. Let us now consider another example problem where the transfer function is full and interactions occur in both directions. Example 3.2.9 Recall the simple mixing tank control problem shown in Fig 3.2 and discussed in Example 3.2.1. If we assume that both the liquid level land tank temperature are measured, then the state variables and output variables are identical. Using Eq. (3.2.28) to convert to the Laplace domain, G(s) = Ct — Ay'B Gs) = Ct - Ay'TAA, ) ‘Thus changing either tank inlet flow rate, u, oF u3, influences both the level and the temperature; hence coupling is in both directions. Erforts have been made to establish measures of interaction in multivariable processes. Perhaps the most widely used measure is the Bristol array [8], which ‘measures the degree of steady-state interaction. To illustrate the application of the Bristol array, consider the 2 x 2 multivariable system shown in Fig. 3.6. The ‘conventional method of tuning the control system would be to first open loop 2 and tune loop 1 so that y, has a good response, then open loop 1 and tune loop 2 until y, has a good response. If there were no interaction (1, = £31 = 0), one could then close both loops and expect the control system to work wel However, in the presence of interactions, the overall control system performance could be quite poor when both loops are closed. Bristol (8] developed a general measure of this interaction in the steady state in a form of the so-called BristolCONTROL OF LUMPED PARAMETER SYSTINS 67 Ay Mee aml An . An ‘The elements of the Bristol array are then the ratio of the steady-state open-loop response and the steady-state closed-loop response when a particular manipu- lated variable is adjusted. It is straightforward to determine the elements of the Bristol array, for the numerator is simply the i, jth element of the open-loop steady-state transfer function, i.e, lim g,(s). The denominator is calculated the closed loops work perfectly, so that y, is constant for k #j. (3) area (3), ne” (Bena However, (2y,/@y,] is just the j, ith clement of the inverse of the steady-state(68 ADVANCED PROCESS CONTROL a Frill, Thus N=[GO] [607], where [],, denotes the i, jth element of the matrix in question. To illustrate the application of the Bristol array, consider the 2 X 2 system shown in Fig. 3.6 with kay ka moo-[it and ‘Several things can be noted from this example which are general properties of the Bristol array: 1. The sum of any row or any column in the Bristol array is unity. 2. When the transfer function matrix is diagonal or is triangular, the Bristol array is the identity matrix. From the definition ofthe interaction measure, itis clear thatthe best possible situation is to have all the off-diagonal terms near zero and the diagonal terms lose 18 unity, ie, 1 o howe]! o 1 ‘This means that there is little interaction and that the closed-loop behavior is similar to the open-loop behavior. As the off-diagonal terms increase in absoluteCONTROL OF LUMPED PARAMETER svSTIMS magnitude andthe diagonal terms depart from 10, more and more interaction is indicated. Positive interaction arses when all elements of A. ate positive: this means that there is some interaction, and one must choose the loop pairings (u, y,) to make the diagonal terms as close to unity a8 possible. ‘Negative interaction occurs when some of the elements of A are negative ‘This means that changing w, in the closed-loop situation has just the opposite effect from changing 1, in the open-loop case-—a potentially dangerous stua- tion. ‘The Bristol array can be used as a guide in choosing the pairing of control and output variables. Some general rules for this pairing are: 1, From examining G(0), pair u, y such that diagonal elements are largest relative to off-diagonal elements. 2. From examining A, pair u, y such that diagonal terms are dominant and close to unity in absolute value. If some diagonal terms are negative, then all must be negative for good performance. Example 3.2.10 Let us consider the pairing of control loops for the distilla- tion column in Example 3.2.8, where 07 0 0 1430 GO)=|20 04 0] GO)'=|-714 25 0 Ba 23 2, 626-274 048 Thus iope0 A -| obo 0 0 10 and the Bristol array predicts no serious steady-state intereactions because the coupling is only in one direction. Tuning can be carried out one loop at 4 time, Le, loop I (u, versus y,) can be tuned, then loop 2 (u, versus y,), and then loop 3 (uy versus y,). This procedure should, in principle, allow reasonable control system performance when all loops are closed. However, as shown in Chap. 6, there ace stil interaction problems with these columns, ‘and more sophisticated multivariable controller designs can be helpful. Example 3.2.11 Let us now consider the stirred tank of Example 32.1 where ‘we pair the control and output variables as shown in Fig. 3.2; ie, the level is controlled by adjusting the hot stream, and the tank tempera- ture by manipulating the cold stream. Recall from Example 3.29 that F= Gawhere y, represents the tank level and y, the tank temperature, Here 1 is the flow rate ofthe hot stream and u, the flow rate of the cold stream. If we pair w; with y, and 1p with y, as shown, itis interesting to calculate the Bristol array to check the interaction. Note that Lf om 0-7 nr, Te-T, roy-!=_ Fle) [%-% -(y- 7) CO = Te Ta) | 2h, 2h, | 0 that the Bristol array becomes (Tx ~ 7) To- Th Teo-T, To~ Ty where we recall that Te, Ty, and T, are the temperatures of the cold inlet stream, the hot inlet stream, and the steady-state exit temperature, respec- tively. Note that since Tq < T, < Ty, all the elements of A are positive, so that only pasitive interactions are possible. Furthermore, if the steady-state operating temperature 7, is close to the hotstream temperature Ty, the Bristol array predicts diagonal elements close to unity, and the present loop pairing will be good. If, on the other hand, the steady-state temperature 7, is close to the cold-stream temperature Tc, then the diagonal elements in the Bristol array are almost zero and the loop pairing should be switched for good perfor- ‘mance. In the event that 7, is midway between T and Tj, then os os 0s 05‘CONTROL OF LUMPED PARAMETER SYSTEMS T1 and neither loop pairing will be very good. Very difficult interactions will be present. This example illustrates the fact that the amount of interaction can depend on the steady-state operating condition in some problems. It should be emphasized that the Bristol array is only a measure of steady-state interactions. Although a great deal of work has been devoted to the study of dynamic interactions, there is, as yet, no generally accepted measure of dynamic interaction comparable to the Bristol array. Such a measure would be useful because itis sometimes possible to have significant dynamic interactions in the absence of important steady-state interactions, Noninteracting Control The performance of a multivariable control system can often be significantly improved by some type of compensation which accounts for interactions. Such improvement is possible even for systems which appear to have only weak interactions according to the Bristol array. There are numerous techniques for designing multivariable feedback controllers with compensation for interactions. One of the classical approaches to the problem is to design a noninteracting controller (9, 10}. To illustrate noninteracting control, let us assume that it is desired to control the process outputs y(1) by adjusting the controls u(t). The feedback control structure, making use of single-loop controllers and a noninter- action compensator G,, may be seen in Fig. 3.7. These single-loop controllers give G, the diagonal form «=| fa, (32.50) The closed-loop transfer function for the noninteracting control scheme shown in Fig. 3.7 is J = (1+ GG,G,.)"'(GG,G.5, + GA) (3281) —- & 20 [connotes Lz, [Compenator]_ [Rona S, sr Figure 3.7 Multivariable controller structure which includes interaction compensator‘Th ADVANCED PROCESS CONTROL, y=T.+ Ta 257) Here G, is a compensator which must be designed to eliminate as much interaction as possible. Ideally G,(s) should be chosen to make T= (1+ 66,6,)"'66,G, (3.282) diagonal and drive T—+I for s = 0 (t+ 20) for some choice of the controller tuning parameters (such as controller gain). Obviously noninteracting control only makes sense when G, G,, G, are square matrices (i, when the number of controls and outputs are the same). Recalling that G, is diagonal, a sufficient condition for T to be diagonal and have T(0) I as the controller gains increase is to require GG, = diag G(s) or G, = G"' diag G(s) (8.2.83) ‘where diag G(s) is a diagonal matrix having the diagonal elements of G(s) along the main diagonal, ie, suls) 0 8x3) diag G(s) = és ° guls) It this compensation were done perfectly, the closed-loop transfer function would take the form H(s) = (I+ GG"! diag GG,)'(6e~' 8 CGF, + GA) 5s) = B()8(5) 5) 1+ gy (s)gu(s) + reas S05) FAQ 28H Notice that there is total decoupling for set-point changes, and that even though ‘each disturbance may influence all the outputs, the effect of the disturbances on any output j, is damped by a single controller g,.. * If there are more controls than outputs, m > then a subset of controls may be chosen to Accomplish decoupling, while if there are more outputs than contol, /'>m, then only partial decoupling is posible.‘CONTROL OF LUMPED PARAMETER SYSTEMS 73 Sometimes it is difficult or impossible to accomplish perfect dynamic compensation of the form of Eq. (3.2.83); however, one may always carry out steady-state decoupling, which eliminates the steady-state interactions. This re- quires lim T(s) in Eq. (3.2.82) to be diagonal. The steady-state compensator is given by G,, = limG,(s) = lim (G(s) diag G(s)] = G;' ding G,, (3.2.85) In this ease the closed-loop transfer functions takes the form J = (1+ GGz! diag G,6,)'(6Gz" diag 6,63, + Gd) (2.2.86) With such steady-state decoupling, the steady-state interactions are eliminated, so that by increasing the controller gains, steady-state offset may more easily be decreased. Even 30, there is still a period of dynamic interaction which could cause single-loop controllers to fight each other if they were too tightly tuned. For the same reasons, significant integral action in the controllers is usually not desired, for this often leads to controllers fighting with each other in an attempt to eliminate offset. Tn our discussion, we have presented noninteracting control via ouput feedback. \t is also possible to use state feedback control to accomplish decou- pling if all the state variables can be measured or inferred (11). ‘We shal illustrate noninteracting controller design with an example prob- lem. Example 3.2.12 Let us consider the distillation column control problem discussed in Example 3.2.8 and design both static and dynamic noninteract- ing controllers for the column. For steady-state decoupling, G,, = G,," diag G,,, where we recall 1+ 9s o i 20. _04 SO-| Tie Te 9 G25) 23 2a Tels Tre Tet 07 0 0 07 0 0 G,=|20 04 0 diagG,=| 0 04 0 | (3.2.88) 23°23 21 0 0 24 and thos 143 0 oO 07 0 0 0 0 G,=|-714 250 0 ||o 04 of=|-214 1 0 626-278 o48|{0 0 21) | 626 -274 1 If we neglect disturbances for the moment, the noninteracting control scheme for the distillation column is shown in Fig. 3.8, where for staticigure 34 Nosinteractng control of a disilation column. decoupling, 1.00 ° ° g,7|-714| e,=| 10 | e,=} 0] (289) 6.26 -2.74 10. ‘Thus u, responds only to errors in y,, while u, responds to errors in both y, and y'y and u; responds to errors in all three output variables. ‘The dynamic compensator design for the column is of the form G, = G(s)" diag G(s) (283) where 149049) ° vi wl 900.4 69 en ; GMs) =| [720 +90 $6907) seo a ry [AG Larue om 19 ~tsol gL 70)] : (+ 10a) (3.2.90) and o7 T+ 9s By io o4 diagGo)-| 0 7S4 0 24 * 2 1+7s‘CONTROL OF LUMPED PARAMETER SySTEMS 75 Hence 1 ° =7.14(1 + 98)(1 + 68) i TF 8s (0) =| [2820+ 990 + anna + 79) ery =27401 + 6NL+ 75) _ =1se(l + 9900479 T¥ ar T+ 10s and the compensator blocks in Fig. 38 take the form 10 (1+ 9901 + 65) ae WT T+ **) 78201 + 95)(1 + 65)(1 + 75) _ gp (1 + 9)(1 + 75) (+8 56+ 10 G291) ° ° 10 ®") ama soyi+7y | * 7 (+85) “ Note that for this example problem dynamic compensation requires dif- ferentiation of the signal coming from the single-loop controllers. The actual performance of steady-state and dynamic noninteracting control designs for this example problem will be described in Chap. 6. Noninteracting controller design does not always work out as well as it did for the previous example. There ean be conditions where implementation is difficult or impossible. For example, if the transfer function G(s) contains time delays, then the dynamic compensator will often contain time leads, requiring a knowledge of the outputs at some future time. Clearly this is impossible to implement exactly. However, in some cases, such a control scheme involving a predictor can yield reasonable results. ‘A problem encountered even more frequently results from the fact that perfect compensation requires a perfect transfer function model. Because the process model is often only approximate, the actual control scheme imple- mented will have the closed-loop transfer function z (1+ G*G"! diag GG)"'G*G"' diag GJ, 3.2.92)116 ADVANCED PROCESS CONTROL ‘where G* is the actual process and G~' is the inverse of the process model. If the differences between the model and the process are too large, the control scheme behaves badly and can even become unstable. ‘These stability problems are most serious if there are right-half-plane zeros in the transfer function. These zeros become unstable poles in G,, and imperfect cancellation of these elements due to imperfections in the model can result in unstable poles in the transfer function. ‘Another potential disadvantage of noninteracting control is that a great deal of the control flexibility is used to achieve noninteraction, sometimes at the expense of overall dynamic response. In cases where some of the interactions gzeatly aid dynamic response, another type of controller design which takes advantage of these beneficial interactions (and only eliminates the most ‘troublesome couplings) might be a better choice. We shall describe some alternative methods in what follows. Set-Point Compensation ‘As an alternative means of eliminating steady-state interactions due to set-point changes, one may use set-point compensation. This scheme, which can be applied either directly to analog controllers by the operator or through supervisory computer control, compensates for steady-state offset due to set-point adjust- ‘ments. Recall that for the control scheme shown in Fig. 33, y= +TA (3.2.57) where G, is a diagonal controller matrix and T=(1+66,)"'6G, (32.58) is, in general, not diagonal. However, suppose Fig. 33 is altered to Fig, 3.9 by Feamo : Fao Figure 39 Multivariable control with set-point compensation,‘CONTROL OF LUMPED PARAMETER svSTEMS 77 the addition of the set-point compensator Ya Shu (3.2.93) where 9, is the actual set point desired and S is the set-point compensation matrix. For example, for the three-input, three-output system in Fig. 3.9, Su Sa Si S=|Su Sa Sy Su Sn Sy Substitution of Eq. (3.2.93) into Eq. (3.2.57) yields F= (I+ GG) '(GG.Sj, + GA) (3.2.94) and $ must be chosen so that T= (1+ 66)"'6G,s (3.2.95) is diagonal and approaches the identity matrix at steady state, Thus by the ultimate value theorem of Laplace transforms, S = lim[( + G6)'66,] (3.2.96) ‘The application of the set-point compensator shall be illustrated by the follow- ing example. Example 3.2.13 Let us consider the distillation column control problem of Example 3.2.12 and apply the set-point compensation scheme shown in Figure 39. If the controller matrix G, consists of three proportional con- ‘rollers kK 0 0 =|]o %& 0 297) 0 0 k, and G,, is given by Eq. (3.2.88), then S=(G6,G)"'(1+ 6,6) = (GG)! +1 (3.2.98) Now (G,G.)"' = G'Gz" where Yk, 0 0 Gt=| 0 1/k, 0 299) 0 0 ik,and et 0) 0) T1425 0 (82.100) 626-274 048 ‘Thus the set-point compensator is o o aoe (82.101) 626-274 08 as fe ee Note that for sufficiently high controller gains (ic. k, —> 20), S-» 1 and no steady-state compensation is necessary. As an exercise, show that the set-point compensator S is the identity matrix Lif PI control is used; ie., if e- afr) (2210 Explain why this must be true. One obvious drawback of the set-point compensator is that it will not improve the response to disturbances because it does not appear in the feedback control loop. However, in making new set-point changes to compensate for sustained disturbances, a set-point compensator will minimize the effects of steady-state interactions. This is especially valuable if one is using supervisory ‘control with local analog controllers in order to minimize the effects of interac- tion. In this case set-point compensation is rather easy to implement. ‘Modal Feedback Control Another approach to multivariable controller design is to use modal feedback control. This technique makes use of the linear nature of the system model to design a control scheme which allows one to specify the closed-loop eigenvalues of the system. To illustrate this technique, let us consider the system described in‘CONTROL OF LIMPED PARAMETER SYSTEMS 79 state space by k= Ax +Bu+ Td (21) yocx (3.2.2) where for this discussion let us assume the number of controls, and the number of outputs, are the same as the number of states. Thus A, B,C are constant nX n matrices, and we shall assume A has real, distinct eigenvalues. These limitations are not crucial to the method (see [12)), but make the explanations to follow easier. Let us further assume a proportional controller on the output, u(!) = -Gy = -G,Cx (32.103) Now let us review the concepts of eigenvectors and eigenvalues. If A is a diagonal matrix of eigenvalues of the m X m matrix » ° A= - . (32.104) ° » then RA=AR (32.105) AL=LA (3.2.106) where R and L are the normalized right and left eigenvectors for the matrix. This means that R, L are the matrices of the n solutions of the equations AVaM Ta LQe ae (2.107) WA=M f= 12... (32.108) where the eigenvalues }, 7 = 1,...,, are solutions of the equations |A-MIl=0 i= 1,2, (32.108) ‘The m vectors I, t, are each divided by a constant to make them orthonormal, Vy dyna 62.110 LR=RL=1 2.11) where (3.2.12){80 ADVANCED PROCESS CONTROL. Thus the properties in Eqs. (3.2.105) and (3.2.106) follow directly from Eqs. (2.107) and G2.111). “Making use of Eqs. (3.2.105) and (3.2.106), one can show that A=LAR (2113) and RAL =A (32414) ‘Thus substituting Eqs. (32.103) and (3.2.114) into Eqs. (3.2.1), one obtains X= (RAL ~ BG.O)x 2.115) Now if we choose the control matrix G, to be G. = BRK 32.116) where K is a diagonal proportional controller matrix k ° ky K- ' (32117) 0 ky and choose the output matrix C = L, then Eq. (3.2.115) becomes %=R(A- K)Lx 32.118) By noting that =Lx 2119) Eq, (3.2118) takes the form y=(A-By (32.120) and clearly (A ~ K) is a diagonal matrix, so the outputs have no interaction and ya behaved 21121) In addition the closed-loop solution to the state equations takes the form x= S axed (2.122) a where @, is a constant determined from the initial conditions. Note that by adjusting k, we ean make the closed-loop eigenvalues as large and negative as we wish, and there is no interaction between the y,. Adjusting the ith controller constant k;, affects only the ith mode. This means we can control the modes of the process, and even though there is state interaction, the output y has no interaction. The disadvantages of this control scheme are that only proportional control is possible, freedom is needed to choose C = L, and experience seems to indicate that tuning isa problem {10}. The block diagram for the control scheme is shown in Fig. 3.10.v4.70 Proportion | [vom |, Outpt Figure 3.10 Block diagram for modal feedback cootrl. In the case where there are more states than controls, it is possible to apply ‘modal control to the first m eigenvectors (where m is the number of control variables) with zero output interaction [12]. Example 3.2.14 Let us illustrate the concepts of modal control through consideration of the simple CSTR of Example 3.2.5, in which the reaction A B—> Cis taking place. It is desired to control the outlet concentration of A and B through manipulation of feed concentrations cyy cay. In dimensionless form (see Example 3.2.5) the modeling equations become ax, Gin - (+ Dada, +m (32.123) ay GE = Dax, ~ (1+ Das)xy + (32.124) where we recall that x, and x, are the dimensionless forms of ¢4, ¢p and uj, ts are dimensionless feed concentrations. The Da, are the Damkdler numbers for each of the reaction steps. Let us suppose that both x, and x, are available as outputs, so any output of the form y= (322) is possible. Now if we apply simple single-loop proportional feedback control on the states (where we assume the problem has been scaled to let x, = 0 be the set point) ay = kus = kaw (32.128) then the system equations are ax Gin - (1+ Day + kun (3.2.126) a P= Dax, ~ (1+ Das + kad (2.127) Even though k,, may be used to control x,, there is a strong influence of x, on the state x3.‘82 ADVANCED PROCESS CONTROL Let us now apply modal control to the problem. We begin by determin- ing the eigenvalues and eigenvectors of the state matrix A, where _[- (+ Da) ° a-[ De | (3.2.128) Thus -(+Dy+) 0 wan =| -a+pysn]-® (3.2.129) yields 2? + (2+ Day + Da)A + (1+ Da)(1+Da,)=0 — (3.2.130) Ay Aa = = (1 + Day), — (1 + Day) (3.2131) ‘Now from standard references on matrix algebra (eg. [1), one obtains the right and left eigenvectors as the nontrivial solutions to (A-XDx=0 (3.2132) which from standard solutions to homogeneous algebraic equations be- ‘comes the nontrivial columns of the adjoint of (A — (1+ Day +2) o waaay) gam so that for M=-(+Da) -( (62:34) d= - (1+ Day) n=[pa,| (32.135) Similarly, the left-hand eigenvectors (or eigenrows) are the nontrivial rows of adj (A — AD; ie, for A= -(+Da) i -[%] (62.136) Da, Da; Now because each eigenvector is uniquely determined only up 10 a multi- plicative constant, one can divide r, and r, by Da\(Da, ~ Day) to make the matrix of right- and lefthand eigenvectors orthonormal [i satisfy Ma -(+Da) (321137) “| Da,Eq.(3.2.111)}. Thus (3.2.138) (3.2139) and one may verify that LR = I. Now if we let C = L and choose G, from Eq, (32.102), then ame 2.140) yam Dayx, + (Da, ~ Das)x, : and the feedback control law becomes w= —RKLx 2.141), where ky ° RKL =| Dayka = kn) 02.14) Bie 4 Day + ee 2.143) ) = (1+ Day + ka)xz (32.144) where one may note that the outputs given by Ea == (1+ Day + kn (3.2145) Bre + Day + kas (3.2.146) show no interact Let us now discuss the features of modal feedback control for this problem. First of all, the outputs (Eq. (3.2.140)] can be controlled indepen- dently without any interaction, and this is an advantage if a meaningful output set point could be devised. However, this is not a problem, because in general yy = Lx, and x, = Ry,; thus one may change back and forth with no difficulties.184 ADVANCED PROCESS CONTROL. Another advantage of the controller matrix (Eq. (3.2.128)] is that if ky and ky, are chosen so that ag ~ ky = Da, — Das (32.147) then Eq. (3.2.144) shows that the state interactions can be eliminated entirely. This is not generally a property of modal control, but is due to the particular structure of this example problem. Because the matrix A must be known and all the states accessible, modal control design in the transform domain is a little artificial. However, one may obviously use it if one desires [10]. For a more complex example of modal control, see the paper by Davison and Chadha [13]. Further Design Techniques ‘There are a whole host of multivariable controller design techniques available (Gee {10} and [14] for an overview). These methods, usually implemented in an interactive mode with a computer, allow the iterative design of the feedback gains until good multivariable controller response is obtained. In addition to ‘optimal control (to be discussed in the next section) and the modal control and noninteractive control design procedures already discussed, other proposed methods include: 1, The commutative-controller technique [10] 2. The inverse Nyquist array technique (10, 14] 3. The characteristic locus technique [10] The reader is urged to consult these references for the details of the design Procedure. A very nice series of case studies showing the performance of some of these designs when applied to the control of a double-ffect evaporator may be found in Ref. [15]. 33 OPTIMAL CONTROL THEORY AND PRACTICE* Another major class of lumped parameter controller design methods involves ‘optimal control. We shall begin our discussion of optimal control theory with the consideration of optimal open-loop control policies. The general class of problems we wish to consider can be represented by the nonlinear modeling equation AO - (9,4) O0 because this is the only policy which minimizes Hand thus satisfies the ‘maximum principle. We note at this juncture that had our objective been to maximize rather than minimize H, then Eq. (3.3.36) would have taken the following form: We if A#C0 Let us now deduce the exact optimal policy. If uy 0 and (0) < 0. This would produce the ‘optimal” policy shown in Fig. 3.125. However, the fact that H must be identically zero along the optimal trajectory when i, is unspecified, leads to a rltat C#K Clearly if \ <0 and A+ C > 0, then tag, <0 which is physically impossi- ble. Therefore, the policy given in Fig. 3.125 cannot be optimal. The policy 4 ue is not optimal because for the temperature 10 increase, tng, > x* ‘must be true; however when u = u,, then both A > 0 and A + C > 0 are satisfied everywhere, and tio, = — 1/(A + C) + Axt/(A + C) and on Oj? It ean be shown [2] that hoosing 84(0) corrected in the gradient direction at each time f produces the feats local improvement in 1. Thus on selecting byl) = (%) e>0 8339) one obtains =f’ S (HY a>o 340) which guarantees 67 > 0 for e small enough that the linear approximation is not violated. ‘These results can be incorporated in the following modified gradient method: 1. Guess (9, 0. 6 < te 2. With this value of (2), integrate the state Eqs. (3.3.1) forward in time to produce X(9), 0S 4 $f 3. With these values of X(1), (9), integrate the adjoint Eqs. (3.3.20) backward in time, 0< 1 <4 4. Correct (+) ‘by Eas (3.3.39) where ¢ is chosen arbitrarily. Evaluate J for this new control w(t) 5. IF T{u(0] > 1{W(9] double e and repeat step 4: If Z[w(2)] <1{6(0), halve e and ‘Tepeat step 4. Do this until a concave function /(e) is formed (see Fig. 3.13). 6. Fit a quadratic 1(e) to these results and predict the optimal value of e, ie‘CONTROL OF LUMPED PARAMETER SYSTEMS 95 ’ 301 Figure 3.13 Determination of the optimal ¢ by a quadratic approximation. 7. Let wo + 2) @34n and return to step 2. 8, Iterate until convergence is attained, Experience has shown that these methods will lead to rapid progress in the first few iterations, but tend to become very slow as the optimum is approached. Even though convergence to the optimum can be proved theoretically, the rate of convergence can be so slow that the exact optimum is never found in a finite number of iterations. For this reason, several second-order methods have been proposed (e., 20-21) which are similar to those for parameter optimization, In addition, conjugate gradient procedures have been developed (22, 23] which showed improved convergence properties over the standard gradient methods. As an illustration of the application of control vector iteration techniques, let us consider an example of chemical reactor control, Example 33.2 Let us consider a batch chemical reactor in which we can control the reaction temperature exactly® and in which we wish to carry out the following reaction: atatiec (33.42) We consider the kinetics and the temperature dependence of the rate constants known, and our objective is to find the optimal temperature control policy, for a fixed batch time, which will maximize the production of the intermediate 2. We note that this is one of the classical batch reactor control problems and that the general scheme given by Eq. (3342) is of considerable practical importance in a number of chemical ‘processing operations, e.g, the oxidation of hydrocarbons or the chlorination of aromatics. In all these cases we may wish to maximize the production of aa intermediate and thus wish to prevent the reaction from going to comple- tion, * tn many practical situations such close temperature contol is, in fact, uit feasible,‘96 ADVANCED PROCESS CONTROL In order to define the problem, let us assume that the reaction is of second order with respect to the first step and of first order with respect to the second step. Thus the material balance on the reacting species may be written as 4 = -K(T) (0) = 1.0 (0343) re KT) kT @(0) =0 (03.44) where e=[A] c= [B] K(T) = Age "8/7 i= 1,2 Let us consider that the temperature is bounded by T. <1) st 6345) The object is to find the open-loop temperature control 7(#) which maxi- mizes the amount of species B present after 1 h of reaction, Thus our objective becomes max[ T= e(0)] (83.46) ‘The additional parameters of the system required to define the problem are given as follows: Ajo = 4000.0 L/ (mol)(s) Ay = 62 X 108/s E, = 5000 cal/ ()(mol) E, = 10,000 cal/ (g)(mol) T, = 298°K T* = 398°K batch time: 1h SouuTion For this problem, the Hamiltonian is = Ay — Adee} ~ Ark Ter (3.3.47) where the adjoint variables A,,A, are given by a a = =A — A(T) AIO 3.3.48) ay, 3H BeBe w= (33.49) and the gradient 84/97 is Fee Gal Mehl EAKMe] 350)‘CONTROL OF LUMPED PARAMETER SYSTEMS 977 ‘The modified gradient algorithm was applied to the problem from two different initial guesses of T(0). The resulting optimal temperature program is shown in Fig. 3.14 together with some of the intermediate iterations. Figure 3.15 shows the optimal yield ¢,(1) as a function of the number of iterations—for two initial guesses of the temperature program. As can be seen, the same optimal temperature program is found in both cases within three to four iterations. An inspection of the graph shows quite clearly that the optimal policy would produce very marked improvements in the yield of the desired intermediate species B. This improvement is found to be ~ 30 percent and 300 percent compared with operation of constant temperatures corresponding to the initial guesses of 398 °K and 298 °K, respectively. This example is thus an illustration of situations in which optimal control may Produce significant improvements in performance. Another approach to the problem of numerically solving the necessary conditions of optimality is to convert the control problem to a two-point boundary-value problem (TPBVP) through the elimination of the control vari- able. These are called indirect or direct substitution methods. The first step in such a procedure is to eliminate the control vector u(t) by solving Eq. (3.3.27) Fint iteration _-Swcon teraton - ss Second iteration sao int terion Init oes “Tne (in) Figure 3.14 The optimal temperature program in Example 33.1,98 ADVANCED PROCESS CONTROL a7 =3982K 03 ‘ t on Se - $1 igure 3:15 Plot of (1) aginst the Number ofeations ‘number of iterations for u(t) explicitly: u(t)= g(x, A) f= 1,2... (3.3.51) tis clear that this may not always be possible; however, it can be done for simple problems. If Eq. (3.3.51) is then substituted into Eqs. (3.3.11) and 3.3.20), the result is set of 2n equations “ =A) x(0) = x (33.52) Banger xy) = (Fi, (333) with split boundary conditions.* This problem, having boundary conditions at ‘two values of the independent variables [a two-point boundary-value problem (TPBYP)}, has a solution which produces the optimal values of x(t), A(i). When these are substituted into Eq. (3.3.51), one obtains the optimal control w(t) What has been effected by the elimination of u(¢) is the trading of a {trajectory optimization problem for a TPBVP. TPBVPs are notoriously difficult * The boundary conditions for Mj) given here assume that all the x(4) are unspecified. In some instances we may end up with two ses of boundary conditions for Eg. (3.3.52) and none for Eg, G33) as will be ilustrated by the example given atthe end ofthe end of the section. However, these problems, to, are two-point boundary-value problems.‘CONTROL OF LUMPED PARAMETER SYSTEMS 99° to solve, even numerically, and thus most of the techniques associated with this approach are techniques for solving TPBVPs. Let us discuss several types of these techniques. One technique for solving TPBVPs is the method of boundary-condition iteration. This approach tries to find, by some iterative procedure, the missing boundary conditions x(1) or AO) so that Eqs. (3.352) and (3.3.53) can be integrated together in the same direction of time. For simple scalar cases a ‘mapping can be done of guessed values of A(0) versus the resulting values of Maae ~ (@G/0x)y as sketched in Fig. 3.16. Obviously this graphical technique will not work well for multivariable problems. However, a number of techniques have been proposed for solving these problems (24 by perturbation methods or by minimizing the error in the boundary conditions by a direct search. There is a basic difficulty with the boundary-condition-iteration approach which often arises in practical problems. The numerical integration of Eqs. (33.52) and (3.3.53) in the same direction is very often unstable. The reason for this behavior is that the state equations are usually stable when integrated forward, but unstable in the reverse direction. Similarly, the adjoint equations are usually unstable when integrated forward, but stable in the reverse direction. This can cause great numerical difficulties which are quite independent of the choice of the proper boundary conditions. Nevertheless, the method has been used successfully for some problems. The reader is urged to refer to standard references (e.., [24 to 26) for further methods of solving TPBVPs. Let us illustrate the direct-substitution-boundary-condition-iteration ap- roach by considering a slight variation on the problem posed in Example 33.1. Example 3.3.3 As before, let us consider the radiant heating of a small billet, as described in Eq. (3.3.29). fect) 10-7 (8329) where 7, is the source temperature, Ty is the initial temperature of the billet, and 7; is the final desired temperature, Our objective is to find the optimal source temperature T,(t) so as to bring the billet to the desired temperature in a minimum ‘time, while ‘minimizing the rate of wear of the refractory roof. Na (0 WO eee Figure 3.16 Solution ofa two-point boundary value problem by a mapping procedure.100 ADVANCED PROCESS CONTROL, ‘The rate of wear may be expressed as Ce (3.3.54) where C; is a constant, ‘Thus the optimality criterion may be written as ‘min. =4 eT, in (1(7,(0)] gras at’) (3.55) where, as before, i is left free and C, denotes the relative value of the roof erosion to operating time. Let us define the variables: mot x=T ujsueTf rece cae xe, and wath ute ret where, a8 before, u* and w, denote the upper and lower limits of the operating temperature, respectively ‘The problem may now be written as min | 1(u) = [1+ Ce* dt (3.3.56) mis) f+ coal] G33 subject 9 Bau-¥ =x xy=% (33.57) Let us follow the procedure set out in Eqs. (3.3.51) to (3.3.53). ‘The Hamiltonian is given as H=XNu- x) + Ce +1 (33.58) Ths, rom Eq, 23.27) we have tat uu 7? (3.3.27) that As cer no ~-¥(-4) eas) On recalling Eq. (3.3.20), ie, by differentiation we obtain Fae (3.3.60)(361) and Eq. (3.3.60). Equations (3.3.60) and (3.3.61) may be solved numerically, eg, by the boundary-condition-iteration technique described above. The result would then be x= x) v=MY) from which u = u(¢) is readily obtained from Eq. (3.359). While in general one would have to iterate to determine the optimal f, from the condition H(d) = 0, in this simple problem Eq. (3.3.58) may be combined with Eq. G.3.59) algebraically to yield Am(-A)-- tere This expression may be solved together with Eq. (3.3.61) from x = x9 to X= x to yield the solution in one integration. We note that the applicability of the direct-substitution technique depends critically on the types of functional relationships involved. Had the rate of wear, Eq, (33.54), been given by an alternative expression, such as Rate of wear = Cje® (33.62) then Hf would have taken the following form: H = Mu — x4) + C exp(u/4) (3.3.63) thus hom ne hcuM expt") 03.64) Clearly, Eq. (3.3.64) cannot be solved explicitly for u. While one could proceed with solving the problem for certain parametric relationships between u and A, this is likely to be cumbersome in the majority of cases. An alternative approach which uses direct search methods is termed control ‘vector parameterization. For open-loop control, one could represent u(t) by a set of trial functions ¢, (1), that is, u(t) = 2 ayoy(t) (3.3.65) fe and use parameter optimization techniques to determine the optimal set of102 ADVANCED PROCESS CONTROL. coefficients a,,. A second approach, which allows a type of optimal closed-loop control, involves generating u, in a feedback form, i.e, by expanding in a set of trial functions of the state variables 4) = 2) by tn Xp) 6.36) land determining the optimal constants ,,. Limited computational experience [16, 27] has shown that this feedback control scheme has much better conver- ‘gence properties than the open-loop version. However, the resulting feedback controller parameters may not be generally optimal; they have been computed for only one initial condition. To be a good set for controller design, they must sive good feedback controller performance for a wide range of initial conditions. Both of these parameterization approaches have the advantages that no adjoint equations need to be solved, and standard parameter optimization techniques such as those discussed in [16] can be used to determine the ‘coefficients. An even greater advantage, as demonstrated in Fig. 3.17, is the ability of the technique to optimize complex process models by allowing a parameter optimization scheme to select the experiments to be performed on this ‘model. This avoids having to modify existing process models in order to perform ‘optimal control calculations—a significant practical advantage. The principal disadvantage of parameterization methods is that the fune- tional form of the optimal control must be specified in advance. This requires ‘much more physical insight than is needed for the previous methods discussed. In the absence of a physical feeling for the general shape of the optimal control, 44 very general functional form [Eq. (3.3.65)] must be used and the optimization performed with respect to a large number of coefficients. On the other hand, if Conte yeor parimetensation = noms, 6) a) octets using Standard proces Sptimiation| ol torn ee a } Et ‘alton 169) I Figare 3.17 The we of control vector para: obiesne Tay seteization with an existing complex ‘action process modelCONTROL OF LUMPED PARAMETER SYSTEMS 103 ‘one has good reason to suspect a particular form of the optimal policy (e.g., a falling temperature profile in Example 3.3.2), then a simple functional form with only a few coefficients should be adequate. A word of caution is in order, however: surprises do arise from optimal control studies (that is why we do them), and in practical problems one should use several types of functional forms to ensure that the functional form chosen is, in fact, general enough. Several other parameterization approaches have been reported and seem to have some merit. See Ref. [27] for a discussion and comparison of these methods. Let us illustrate this method with an example problem.* Example 3.3.4 Consider the problem of determining the optimal start-up control scheme for a nonlinear CSTR in which the exothermic, first-order reaction A —» B is taking place. The modeling equations take the form My a OF Zoli = CAE) F/I 5 a7 96G- T+ 7G Kee 'e-au(T-T.) O0 4() = {= vaca (3.3.8) This behavior, plotted in Fig. 3.20, is called a bang-bang control policy. The points), f, where h,(¢) changes sign are called switching times. There is another special situation which occurs when h, = 0 over some interval of time. An examination of Fig. 3.21 shows us that h, = 0 over 1, $1 $f Since this causes the control to vanish from # in that interval, itis not clear how we can determine the optimal value of u(#) over f <1 < fy. The control over this interval is called a singular are, and these especially difficult problems are called singular control problems. Whenever one encounters a problem such that the Hessian matrix an at 3.3.6) ae 33.69) is singular over some interval f, <1 0. For values of too small, u(t) can become negative, violating this constraint, while for too large, little dynamic control action is possible because K((1), K3(0) remain close to zero. Thus for good optimal linear-quadratic controller design, one must tune the weighting parameters (such as @ and f here) in order to obtain the desired controller response. From the previous example one may note that the linear-quadratic formula- tion produces a proportional state feedback controller. From classical controlCONTROL OF LUMPED PARAMETER SYSTEMS 113 theory one recognizes that proportional controllers lead to offset when there are set-point changes or load changes in the process; thus, it would be desirable to formulate the optimal feedback control problem so as to allow integral control action which would eliminate these offsets. There are several possible ways of doing this. One method of including integral action is to include du/dt terms in the ‘objective functional so that Eq. (3.3.71) becomes T= tS), +5 [TP + wT eA) a (3.392) where the state equation (3.3.70) must be differentiated to yield (for constant matrices A, B) 8 = At + Ba ‘Then the problem may be reformulated by letting, Woe mex mee wef (33.93) so that 0294) i x + and the objective becomes h., BLURS ve « om By applying the feedback control law (Eq. (3.3.82)] to this reformulated problem, we obtain ] = 396) v= -K(Ow = =[K,) K, which, upon integration, takes the form u(t) = Kx(9) — Sits, ~ K,)x(s) dt 397) Which has somewhat complicate integral action, If one allows ,» co, then Ky and K, are constants and the feedback control law is u(t) = —K,x(0) — Ky Six (3.3.98) an “optimal” proportional-integral contoller:* One may justifiably question the physical meaning of minimizing the time derivative of the control action in Eq. (3.3.92) and how this relates to reducing, offset in the state variables. At the moment this relation is unclear. * Through a slightly diferent transformation of state, derivative action may be obtained as well, (see 0p.114 ADVANCED PROCESS CONTROL ‘A second means of incorporating integral action into the controller [31, 32] is to augment the state variables to include p new variables 2(1) where a=cx (33.99) are those state variables for which integral action is desired. Thus the new state & of dimension n + p is 2-[3] (63.100) When the objective functional is also modified to accommodate the new state variable, ie., T= IS +5 [Rh +0" Ea) di (33.101) the linear-quadratie optimal control law takes the form um WK = —Kyx ~ Ky = — Kx — K,C* fix at (3.3.102) which naturally includes integral action. Note that necessarily p < m, that is, the number of state variables for which integral action is desired cannot be larger than the number of control variables. Another derivation of proportional and integral “optimal” feedback control is given in Sec. 54, where stochastic contol is discussed. There integral action arises naturally as a means of dealing with random process disturbances. Optimal Linear-Quadratic Feedback Control of Nonlinear Systems tis possible to extend the results of the linear-quadratic problem to nonlinear systems so as to produce an optimal feedback control law. Let us consider the nonlinear optimal control problem given by the modeling equations Satey = (33.103) tnd the contol objective fumetional TLu(s)] = G(x(y)) + Sir u) de (33.104) Now if one expands the objective functional to second-order about a nominal control and state trajectory (1), X(1) while adjoining the state constraints, the variation in 7 is a1 (32), oxy) + 7023) OKO HY) an #rreo ond + f°) 90+ 00" EE) ou (GE) oe + Your(22) oc or( SL) we BE oe] 09 2 ax ou dt‘CONTROL OF LUMPED PARAMETER SYSTEMS 115 where it is useful to recall that éxex-X dusu-a and H=F+.t Now let us suppose our nominal trajectory U(0), X(é) satisfies the first-order necessary conditions for open-loop optimality, ie., aH ant aH) 77, - (26 May A.-M) wwy-(25) cr ‘where we assume x(‘q) is specified (x(q) = X) and x(4) is unspecified. Thus X(t), {i(1) are the open-loop optimal controls for a particular initial condition x(‘q) = Ko, In this case Eq. (3.3.105) has only the remaining terms: ar ff Soe( 22) 0 (Pou + Loo +$oru( ZZ), ants 03.107) Now let us note that if we change the initial condition x(1.), the entire nonlinear open-loop optimal control policy must be recalculated, because unlike linear problems, in which superposition may be used to quickly adjust for changes in initial conditions, nonlinear optimal control problems depend nonlin- carly on the initial conditions. However, itis possible to develop a perturbation approach to solve this problem. If we consider the system equations linearized about the open-loop optimal policy X(1), u(:) for a fixed initial condition 2Xx(ig) = py We obtain the perturbation equations AEH) (an (Jan ant) 03100 hich repost he sem avr orn codons in sme domain ove to the nominal value %, ie., for |Sx(%)| TC) i tree different ways: (@) Apply the maximum principle in order to calculate the optimal control using a gradient procedure, Determine the optimal program from two initial guesses of (0). wt) = a + aye + aye? ultivariable search routine (se Ref. [16) o determine the optimal valves of ay a, and oy. Usea Compare your results with thse found by the gradient method. (©) Determine an optimal feedback control law of the form. u(x) = bo + bux + baat by using a multivariable search routine to determine the optimal contoller parameters by by, bs ‘Compare with the results of parts (a) and (0). ‘32 A simple suspension system is to be designed for a high-speed rapid transit veicle, A simple laboratory model of the system i sketched inthe figure below, where springs (pring constant K’) are used to cushion the vertical motions in the absence offeedhack contol. However, itis desired to esig a feedback controller to improve the stability and performance of the suspension sytem. To accomplish this, a force (1) may be applied to the bouncing system.130 ADVANCED PROCESS CONTROL ‘The modeling equations are “e wie Kees oew where @ is the force of gravity and Wis the maze ofthe system. By noting that at steady state wan 0 K+ ‘we may define normalized deviation variables and parameters neiey neh we ek to yi the model és a7" “ Sonny se Based on this information: (e)Make se ofthe results in Sec. 3.3 to find the optimal linear quadratic feedback control aw which minimizes 1m {GR + aad + But) de (Cary out the computation for a= 1, = 025, K = 3,4 = 20, x(0) = 1.0, x40) = 10. (@) The waditional feedback controller for suspension systems isthe shock absorber, which has the form w= — Ci, (pure derivative action). Compare your “optimal controler” with thie tradi ‘ional one. ‘33 Cay out the computations to determine the near-optimal inear quadeatic feedback controller for the problem discussed in Example 337. Use the following parameters: « = 1000, 7 = 10, y= 13,30) = 15, 4 = 20, and a = 0.25. After the optimal open-loop policy has been fond for MO) ="I5, determine the near-optimal feedback controler performance for (0) = 14 and s(0) = 1L6. How does the controller perform for this highly nonlinear system? 34 Consider the tank network shown in the figure below. Liquid flows from one tank tothe next, ‘ith the outfow of each tank assumed tobe proportional to the level in that tank. The outlet ofthe second tank is split, witha fraction F exiting and the remainder pumped back to the fist tank. The ‘modeling equations can be assumed as a Bina Oh a Bn ah (B+ It is desired to control the ligui levels in these tanks by choosing a controlled inlet flow position (Gither u of uy) a8 shown. As a means of aiding your choice, (o) Determine ifthe controllabisty conditions are satisfied for my alone (valve wt constant flow ate) and for wy along (valve mat constant flow rat) (@) Are there other practical conditions to be considered which are not treated by formal controllability condition, but which may influence your choice? Discus these. () If the recyle loop were removed, modify the modeling equations, and reevaluate contola- bility asin part (2).CONTROL OF LUMPED PARAMETER SYSTEMS 131 35 For the vehicle suspension problem discused in Prob. 32, develop a transform-domain repre- ‘entation between the measured vertical height y = xy and control fore wie, (9) = (9800) ‘Determine (3) Hint: See Eq. (3.350) and the accompanying discussion] [36 You have been given the multivariable control system. Tem Tea ce - y _ Te Fe 4) =| in the Laplace domain, Convert this to an equivalent set of ordinary differential equations of the form * a act Bw yn ee [47 Consider the water tank system of Prob. 34 with both control valves uy, available. Carry out the following types of controller designs: (@) Two single-loop proportional controller. () Modal feedback control (©) Set-point compensation for the controler in part (a) (d) Complete dynamic noninteracting control. Compare the contller responses (interaction, offset, ete) and the relative ease of implementation of the controllers. ‘As test disturbances, consider a unit step change inthe hy setpoint while endeavoring to keep ‘ny constant. Conversely, considera unit step change inf while hy i held constant. Make use of the following parameters: a = 3, B= 2, y= 2. For each design, choose the more convenient system representation (ie, time domain or transform domain)132. ADVANCED PROCESS CONTROL, ‘34 Pulse testing of « distilaion column yields the following dynamic model between product ‘concentrations y,, 72 and product drawolf rates 1 30) = G(s)a(9) where 1. 208 Wy “ Tie Fe | cs Os) ioe 405) » 4 Te T+ (@) Convert the above transform state space model to time-domain model inthe form x= A+ Bu yoo Is this «minimal realization in state space? (@) Is the distilation column ouput controllable! Why? (©) Write down the block diagram and the design equations fora noninteracting proportional feedback controller which eliminates steady-state interactions. What is the closed-loop transfer function between y and set pointy? 39 Extend the linear quadratic optimal feedback control law of Sec. 33 to include opuimal feedforward contol when measured disturbances are included in the model Le, find the optimal ‘eedback feedforward control law forthe system a Boat But Dd xi) = 6 where D isan n X k matrix and di a vector of disturbances. (Hint: see Ref. (25D. ‘340 Consider the nonisothermal CSTR in which the exothermic reaction A +B is being carried ‘ou. The modeling equations in dimensionless form may be writen x 2 She ny + Da xe ty sy Sho (1+ Bye + B Dall ~ xe + Buy (2) Determine if this system is controllable when operated around the steady sate Xin x, resulting from uj = ty, = 0. (8) Design & proportional feedback control system which contains a ing compensator. Bepn by liearzation around the steady-state noted i (©) Simulate th reactor forthe parameters Da = 0.1, 8 = 30, B = 19, and demonstrate your controller performance when applied fo the nonlinear system. ‘3.11 Consider the CSTR described in Example 3.25 with continuous modeling equations (+ Dads, +m Daye, ~ (1+ Dade + Its planned to implement P + 1 control relating uj to x1g~ x, and wy to 4 ~ x, under DDC ‘mode with sampling time Ar. Here xq, x, ae the state st points. Convert the continuous model 0 the equivalent discrete model necessary for DDC design. Draw the block diagram and specify final controller design equations‘CHAPTER FOUR CONTROL OF DISTRIBUTED PARAMETER SYSTEMS 4.1 INTRODUCTION Distributed parameter systems are distinguished by the fact that the states, controls, and outputs may depend on spatial position. Thus the natural form of the system model is the partial differential equation, integral equation, or transcendental transfer function. One particularly important class of distributed parameter systems consists of those having pure time delays. There exists a wide range of industrially important distributed parameter control problems (see [1-4] for a selection); however, we shall choose two simple example problems to illustrate some of the fundamental concepts. Example 4.1.1° Consider the problem of reheating a steel slab by thermal radiation (for rolling) in a batch furnace as sketched in Fig. 4.1. For proper rolling characteristics, it is necessary for the slab to have a specified temperature distribution 7,(2). Thus our problem is to control the heat flux to the surface of the slab in such a way as to approach this desired temperature distribution in some optimal fashion. * This examples taken from (5} and reprinted by permission of John Wiley and Sons, ne. 13134 ADVANCED PROCESS CONTROL TP rere 41 Radiant heating of «slab To be more precise, let us consider the modeling equations for the slab arte.) __ 1 Af a(7) aM, )/d2] sz

You might also like

• Process Control-Thomas E Marlin

  

  73% (11)

  Process Control-Thomas E Marlin

  1,026 pages
• Process Control Modeling Design and Simulation B Wayne Bequettepdf PDF Free

  

  100% (2)

  Process Control Modeling Design and Simulation B Wayne Bequettepdf PDF Free

  933 pages
• B. Wayne Bequette - Process Control - Modeling, Design and Simulation - Prentice Hall (2003)

  

  No ratings yet

  B. Wayne Bequette - Process Control - Modeling, Design and Simulation - Prentice Hall (2003)

  1,058 pages
• Process Control Systems - Shinskey - 4th

  

  100% (4)

  Process Control Systems - Shinskey - 4th

  229 pages
• Practical Chemical Process Optimization Book

  

  No ratings yet

  Practical Chemical Process Optimization Book

  444 pages
• Fundamentals of Process Control Theory ThirdEd - Murrill - TOC

  

  100% (1)

  Fundamentals of Process Control Theory ThirdEd - Murrill - TOC

  4 pages
• Plantwide Dynamic Simulators Luyben 2002

  

  100% (2)

  Plantwide Dynamic Simulators Luyben 2002

  448 pages
• Chemical Reactor Analysis and Design Rawlings PDF

  

  No ratings yet

  Chemical Reactor Analysis and Design Rawlings PDF

  319 pages
• Principles and Practice Control Process 3rd Edtion Smitih Corripio

  

  50% (2)

  Principles and Practice Control Process 3rd Edtion Smitih Corripio

  10 pages
• Simple Dew Point Control With Propane Loop For Chilling (HYSYS V8.6)

  

  100% (1)

  Simple Dew Point Control With Propane Loop For Chilling (HYSYS V8.6)

  32 pages
• CPB 30004 - CHAPTER 1 - INTRO Sept 2014

  

  No ratings yet

  CPB 30004 - CHAPTER 1 - INTRO Sept 2014

  52 pages
• Chemical Process Control - Stephanopoulos PDF

  

  83% (53)

  Chemical Process Control - Stephanopoulos PDF

  376 pages
• Marlin - Process Control (Searchable)

  

  100% (5)

  Marlin - Process Control (Searchable)

  1,023 pages
• Process Control PETER HARRIOTT.

  

  100% (4)

  Process Control PETER HARRIOTT.

  396 pages
• Instrumentation and Control of Heat Exchanger

  

  No ratings yet

  Instrumentation and Control of Heat Exchanger

  29 pages
• Advanced Process Control

  

  0% (1)

  Advanced Process Control

  2 pages
• Distillation Column Control Design

  

  0% (1)

  Distillation Column Control Design

  22 pages
• Advanced Process Control-HYSYS

  

  No ratings yet

  Advanced Process Control-HYSYS

  27 pages
• Chemical Process Control - George Stephanopoulos (Prentice Hall, 1984)

  

  50% (2)

  Chemical Process Control - George Stephanopoulos (Prentice Hall, 1984)

  369 pages
• Transport Phenomena 2nd Edition (0471999903)

  

  63% (8)

  Transport Phenomena 2nd Edition (0471999903)

  303 pages
• Nitin Kaistha - Plantwide Control of Chemical Processes-NPTEL

  

  No ratings yet

  Nitin Kaistha - Plantwide Control of Chemical Processes-NPTEL

  223 pages
• Process Dynamics and Control, 1st Edition

  

  No ratings yet

  Process Dynamics and Control, 1st Edition

  735 pages
• Chemical Reactor CSTR

  

  No ratings yet

  Chemical Reactor CSTR

  88 pages
• Mathematical Modelling of Chemical Processes

  

  No ratings yet

  Mathematical Modelling of Chemical Processes

  39 pages
• Introduction To Process Control

  

  No ratings yet

  Introduction To Process Control

  19 pages
• Process-Dynamics Modeling Analysis and Simulation Wayne Bequette PDF

  

  No ratings yet

  Process-Dynamics Modeling Analysis and Simulation Wayne Bequette PDF

  632 pages
• Editorial Choosing Advanced Control

  

  No ratings yet

  Editorial Choosing Advanced Control

  9 pages
• HiDC PHD Thesis

  

  No ratings yet

  HiDC PHD Thesis

  225 pages
• Design To Distillation Control System

  

  100% (2)

  Design To Distillation Control System

  529 pages
• Advanced Control Unleashed

  

  100% (1)

  Advanced Control Unleashed

  452 pages
• Control System Design

  

  100% (1)

  Control System Design

  26 pages
• Intro To Process Control

  

  100% (1)

  Intro To Process Control

  19 pages
• Process Dynamics and Control Notes PDF

  

  100% (3)

  Process Dynamics and Control Notes PDF

  81 pages
• UniSim Design OLI Interface Reference Guide

  

  No ratings yet

  UniSim Design OLI Interface Reference Guide

  232 pages
• Aske Thesis Draft 130908

  

  No ratings yet

  Aske Thesis Draft 130908

  182 pages
• ECUST PROII Advanced Training

  

  100% (2)

  ECUST PROII Advanced Training

  118 pages
• Chapter 3 - Batch Processes

  

  100% (1)

  Chapter 3 - Batch Processes

  54 pages
• 1 Process Dynamics and Control No1

  

  100% (2)

  1 Process Dynamics and Control No1

  32 pages
• Process Flowsheeting by A. W. Westerberg, H. P. Hutchison, R. L. Motard, P. Winter PDF

  

  No ratings yet

  Process Flowsheeting by A. W. Westerberg, H. P. Hutchison, R. L. Motard, P. Winter PDF

  131 pages
• Introduction To Chemical Process Simulators

  

  No ratings yet

  Introduction To Chemical Process Simulators

  12 pages
• Process Systems Analysis and Control - S. E. LeBlanc and D. R. Coughanowr

  

  No ratings yet

  Process Systems Analysis and Control - S. E. LeBlanc and D. R. Coughanowr

  63 pages
• Process Control: Modeling, Design, and Simulation 2nd Edition B. Wayne Bequette download pdf

  

  100% (1)

  Process Control: Modeling, Design, and Simulation 2nd Edition B. Wayne Bequette download pdf

  51 pages
• Process Dynamics and Control 3rd Edition Chapter 12 Problem 3E Solution

  

  0% (1)

  Process Dynamics and Control 3rd Edition Chapter 12 Problem 3E Solution

  8 pages
• Process Plant Instrumentation Design and Upgrade

  

  100% (1)

  Process Plant Instrumentation Design and Upgrade

  364 pages
• Notes On Chemical Process Control and Chemical Modeling Stephanopulos

  

  100% (5)

  Notes On Chemical Process Control and Chemical Modeling Stephanopulos

  667 pages
• Process Reaction Curve

  

  100% (1)

  Process Reaction Curve

  6 pages
• Group-1 Introduction ChE-5202

  

  No ratings yet

  Group-1 Introduction ChE-5202

  5 pages
• Process Dynamics & Control: BSC, Chemical Engineering 8 Semester

  

  No ratings yet

  Process Dynamics & Control: BSC, Chemical Engineering 8 Semester

  72 pages
• Process Control Engineering - A Textbook For Chemical, Mechanical and Electrical Engineers, 1st Edition

  

  100% (1)

  Process Control Engineering - A Textbook For Chemical, Mechanical and Electrical Engineers, 1st Edition

  423 pages
• McGraw Hill - Process Control Systems (Shinskey)

  

  0% (2)

  McGraw Hill - Process Control Systems (Shinskey)

  8 pages
• Marlin EntireBook

  

  100% (1)

  Marlin EntireBook

  1,008 pages
• 4-uom4dvvr

  

  No ratings yet

  4-uom4dvvr

  6 pages
• Control

  

  No ratings yet

  Control

  43 pages
• Control Engineering For Chemical Engineers

  

  No ratings yet

  Control Engineering For Chemical Engineers

  11 pages
• Ce-307 - Chemical Engineering Design

  

  No ratings yet

  Ce-307 - Chemical Engineering Design

  41 pages
• Practical Process Control

  

  100% (3)

  Practical Process Control

  325 pages
• Process Control

  

  No ratings yet

  Process Control

  9 pages
• Download ebooks file (Ebook) Applied Process Control: Essential Methods by Michael Mulholland ISBN 9783527341191, 3527341196 all chapters

  

  100% (3)

  Download ebooks file (Ebook) Applied Process Control: Essential Methods by Michael Mulholland ISBN 9783527341191, 3527341196 all chapters

  81 pages
• Process Control in Chemical Industry

  

  No ratings yet

  Process Control in Chemical Industry

  2 pages
• 5-bru1c2fa

  

  No ratings yet

  5-bru1c2fa

  6 pages