VARIABLE REGRESSION ESTIMATION OF UNKNOWN DELAY By Ashraf Elnaggar B. Sc. Electrical Engineering, Cairo University, Egypt, 1982 M . Sc. Electrical Engineering, Cairo University, Egypt, 1985 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE STUDIES ELECTRICAL ENGINEERING We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA December 1990 © Ashraf Elnaggar, 1990 SYSTEM In presenting degree freely this at the thesis in partial fulfilment of University of British Columbia, I agree available for copying of department publication this or of reference thesis by this for his and study. scholarly or thesis for her I further purposes representatives. the of may be It is financial gain shall not £l£Cr£lCr!fL £A/6ZA/£££z/U6> The University of British Columbia Vancouver, Canada Date DE-6 (2/88) 6zt>y 4 ,< tyft that agree permission. Department requirements that the Library an advanced shall make it permission for granted by understood be for allowed the that without extensive head of my copying or my written Abstract This thesis describes a novel approach to model and estimate systems of unknown delay. The a-priori knowledge available about the systems is fully utilized so that the number of parameters to be estimated equals the number of unknowns in the systems. Existing methods represent the single unknown system delay by a large number of unknown parameters in the system model. The purpose of this thesis is to develop new methods of modelling the systems so that the unknowns are directly estimated. The Variable Regression Estimation technique is developed to provide direct delay estimation. The delay estimation requires minimum excitation and is robust, bounded, and it converges to the true value for first-order and second-order systems. The delay estimation provides a good model approximation for high-order systems and the model is always stable and matches the frequency response of the system at any given frequency. The new delay estimation method is coupled with the Pole Placement, Dahlin and the Generalized Predictive Controller ( G P C ) design and adaptive versions of these controllers result. The new adaptive G P C has the same closed-loop performance for different values of system delay. This was not achievable in the original adaptive G P C . The adaptive controllers with direct delay estimation can regulate systems with dominant time delay with minimum parameters in the controller and the system model. The delay does not lose identifiability in closed-loop estimation. Experiments on the delay estimation show excellent agreement with the theoretical analysis of the proposed methods. ii Acknowledgment I would like to express my deep thanks and gratitude to my research supervisor Dr. G . A . D u m o n t for his guidance and supervision throughout my study. I'd like also to thank him and the Pulp A n d Paper Research Institute of Canada ( P A P R I C A N ) for providing the financial support and computer facilities since the beginning of my study. I'd like also to thank Dr. W . G . D u n f o r d who accepted me first in U B C and provided some financial assistance. The basic idea of Variable Regression Estimation was born when I was taking E E 570 course and I thank Dr. M . S. Dayies, the course instructor, for referring to the problem of systems of unknown delay in his course and for encouraging me to study this problem in depth. Every idea, problem and result in my thesis was discussed with my colleague A b d e l L a t i f Elshafei and many improvements of the basic V R E are the results of our continuing discussions. He has also read the different versions of the thesis and provided me with valuable suggestions. At the end of my student life, I thank all who have helped me through it. Special thanks and gratitude to my late father and to my mother who have provided me with good life and excellent studying environment for many years and to my brother A y m a n who have always believed in me. I also thank my teachers at the different schools and universities in Egypt who encouraged and helped me. To my little, girl M i r i a m I give my thanks for the blessing and joy that came to my life with her and for making me believe more in adaptive control. Finally, my deep gratitude and thanks to my wife M o n a for the unlimited support and encouragement in she gives me and for the lovely home life that we share. IV Table of Contents Abstract ii Acknowledgment iii List of F i g u r e s 1 2 Introduction 1 1.1 Background 1 1.2 Motivation for the Research Topic 2 1.3 Main Contributions of the Thesis 3 1.4 Thesis Outline 5 I n t r o d u c t i o n to D e l a y E s t i m a t i o n 2.1 2.2 3 viii 7 Methods of Delay Estimation 10 2.1.1 The Extended Numerator Polynomial Method 11 2.1.2 The Rational Approximation Method 12 2.1.3 The Cross-Correlation Method 13 Conclusions 14 D e l a y E s t i m a t i o n U s i n g V a r i a b l e Regression 15 3.1 Variable Regression 15 3.2 The Estimated Model Variable Regression Estimation ( E M V R E ) 3.2.1 3.3 . . . . Advantages and disadvantages of the E M V R E Conclusions . . . v 22 29 . 32 4 5 T h e F i x e d M o d e l V a r i a b l e Regression E s t i m a t i o n ( F M V R E ) 33 4.1 The Fixed Model Parameters 38 4.2 Estimation of the System Parameters 40 4.3 First-Order Systems 42 4.4 Second-Order Systems 4.5 Higher-Order Systems 47 4.6 Effect of Random Noise 50 4.7 Excitation Requirement for F M V R E 53 4.7.1 55 Sinusoidal Excitation 4.8 Robustness of the F M V R E 4.9 Systems with fractional delay 43 58 .' 67 4.10 Conclusions 72 V R E - B a s e d Adaptive Control 73 5.1 Adaptive Pole-Placement 76 5.2 Adaptive Dahlin Controller 80 5.3 Adaptive Generalized Predictive controller 90 5.4 Delay Estimation in Closed-Loop 94 5.4.1 Deterministic Systems 94 5.4.2 Stochastic Systems 99 5.5 6 • • • • Conclusion 102 E x p e r i m e n t s on D e l a y E s t i m a t i o n 104 6.1 Temperature Control Experiment 104 6.2 Reject Pulp Refiner Experiment 108 6.3 Conclusions 110 VI 7 Conclusions 114 7.1 Summary of the Work 114 7.2 Future Research Directions 116 Bibliography 117 vii List of Figures 3.1 Minimization of J 21 3.2 E M V R E of a first order system 27 3.3 Estimation and adaptive P-P control using the E M V R E 28 3.4 E M V R E at different sampling rates 30 3.5 E M V R E of a high order system 31 4.6 Impulse response of G\ 37 4.7 Delay estimation using Ziegler-Nichols and F M V R E 41 4.8 F M V R E of a first order system 4.9 Minimum T, required for a second order system . . 44 46 4.10 F M V R E of a second order system 48 4.11 Response of a high order system 49 4.12 F M V R E of a high-order system 51 4.13 F M V R E in presence of coloured noise 54 4.14 F M V R E using only one sinusoid 59 4.15 Bode plot of the third-order system 63 4.16 Delay estimation for the case of model order mismatch 65 4.17 Parameter estimation for the case of model order mismatch 66 4.18 Delay estimation for system with fractional delay 70 4.19 Parameter estimation for system with fractional delay 71 5.20 Structure of the Estimators and the adaptive controller 5.21 Open-loop response and F M V R E delay estimation viii . . . . . . . . . 74 77 5.22 Adaptive P P / F M V R E response and delay estimation 81 5.23 Adaptive P P / F M V R E parameter estimates 82 5.24 Adaptive PP/extended polynomial response 83 5.25 Adaptive P P / E M V R E response and delay estimation 84 5.26 Adaptive D a h l i n / F M V R E response and delay estimation 86 5.27 Adaptive D a h l i n / F M V R E parameter estimates 87 5.28 Adaptive Dahlin/extended polynomial response 88 5.29 Adaptive D a h l i n / E M V R E response and delay estimation 89 5.30 Adaptive G P C / F M V R E response and delay estimation 95 5.31 Adaptive G P C / F M V R E parameter estimates 96 5.32 Adaptive GPC/extended polynomial response 97 5.33 Adaptive G P C / E M V R E response and delay estimation 98 5.34 Correlation functions of the process noise 101 6.35 Temperature control experiment 105 6.36 Delay estimation and temperature variation 107 6.37 Schematic representation of a pulp refiner 109 6.38 Measurements on the pulp refiner Ill 6.39 Delay and gain estimation of the pulp refiner 112 6.40 Correlation functions of the pulp refiner measurements 113 ix Chapter 1 Introduction 1.1 Background The field of parameter estimation is of great importance in control applications where it is generally difficult, if not impossible, to develop an exact model of the process to be controlled. Many processes are so complex that developing an exact model from the basic principles is very difficult. Even when an exact model can be derived, the parameters of this model may not be known or may change with time, temperature, operating point, etc. in an unknown way. For example, the resistance of an electric motor depends on its temperature, which is not usually measurable. For these sys- tems the parameter estimation methods offers a simple way of identifying the process parameters using input-output measurements. Parameter estimation is a very active research area in control, signal processing and statistics literature. The introduction and successful use of the adaptive techniques for automatic control and signal processing over the last 30 years has encouraged research in estimation methods and led to important developments of the parameter estimation methods and the analysis of their properties. The least-squares method is the most important method of parameter estimation and is heavily used in practice. The basic idea can be traced back to Gauss (1809). The recursive form of least-squares has apparently been found by several authors. The original reference seems to be Plackett (1950). There exist many papers dealing with 1 2 Chapter 1. Introduction different aspects of the least-squares method. Studies of the consistency properties and the implementation aspects can be found in many papers and books, for example Ljung and Soderstrom (1987) and Goodwin and Sin (1984). In this thesis, least-squares estimation methods are applied in new ways to important control problems. Detailed surveys will be given at the time of introducing those problems. 1.2 M o t i v a t i o n for the Research Topic The standard estimation algorithms have already been applied to thousands of systems. Although many systems can be systematically modelled by a linear input-output model, some a-priori information about the system is always required to limit the time and effort spent in the modelling process. There is an infinite number of ways that describe the input-output relations of the system. For example a first-order Auto-Regressive Moving-Average ( A R M A ) model can be represented by an equivalent infinite order Moving-Average (MA) model. In this case, the A R M A model has only two unknowns (the pole location and the gain), while the equivalent M A model has an infinite number of unknowns (the impulse response coefficients). There is a direct correlation between the number of unknown parameters in the model and the signal excitation required to estimate those parameters correctly. Also, the rate of convergence of the estimation process decreases rapidly when the number of parameters increases. For the first-order example, it is the a-priori knowledge of the system order that makes it possible to represent the system by only two parameters in the A R M A model instead of thirty or more parameters in the M A model. It is surprising to discover that many adaptive control applications do not utilize some crucial information available about the systems they control. In this thesis, an 3 Chapter 1. Introduction important case of adaptive control is studied and the a-priori available system information is fully utilized. A minimum number of parameters are estimated to represent the system. Accurate, fast and easy estimation of models results from the full use of the available information. The adaptive control area covered in this thesis is: Systems with delay. These systems are very common in industry. Conveyor belts, long pipe-lines or any transportation line are typical sources of system delay where the system input goes to one end of the transportation line (e. g. output valve of a reservoir tank) and the output is measured at the other end of the line (e. g. the level of receiving tank). Traditional on-line methods of delay estimation do not directly estimate the system delay but rather replace it by a large number of unknown parameters in the system model. In this thesis, the unknown delay is directly estimated and no additional parameters are used in the system model. 1.3 M a i n C o n t r i b u t i o n s of the Thesis A novel method of on-line delay estimation using the new concept of Variable Regression Estimation ( V R E ) is developed and demonstrated analytically, on computer simulations and experimentally. • The variable regression can be added to any standard parameter estimation with minimum additional computations. The number of estimated parameters is the same as that in the known delay case. This means faster convergence and less excitation required. In the case of known process parameters, any non-constant signal is enough to estimate the delay. • For systems with a small range of parameter variations, the Estimated Model V R E ( E M V R E ) is introduced. Both the system parameters and delay are estimated interactively. The delay estimation is unbiased even if the parameter er 1. Introduction 4 estimates are biased due to the existence of coloured noise. For systems of large or unknown parameter variations, the Fixed Model V R E ( F M V R E ) is introduced. The delay estimation is independent of the parameter estimation. The delay estimation is unbiased for first and second-order systems and for higher-order systems of pole excess equal to one. For other systems, the delay estimation converges to a value close to the Ziegler-Nichols approximation. The F M V R E delay estimate is bounded and unbiased in the presence of coloured noise. The excitation richness required for delay estimation alone is less than that required to estimate the parameters if the delay is known. Excitation richness as low as one sinusoid can be enough to estimate the delay correctly. The estimated model is stable if the system is stable even in the presence of model order mismatch. If fractional delay exists, the F M V R E may have a bias of one in the delay estimation. This bias can easily be corrected by extending the model numerator polynomial by one. Unlike the model parameters, the model delay cannot lose identifiability in closedloop. The noise has no effect on the delay estimation in open-loop identification. In closed-loop identification, a delay estimation bias may result because of the noise. To avoid this bias, different options are given. V R E is linked with controller designs and the adaptive versions of pole placement, Dahlin and generalized predictive controllers successfully applied to systems with 5 Chapter 1. Introduction variable parameters and delay. The VRE-based G P C provides the same closedloop performance over the range of delay variation. This was not possible in the original G P C . • Experiments for closed-loop and open-loop delay estimation are performed. Adaptive D a h l i n / F M V R E is used to control the temperature in an experiment having long delay. The results show fast adaptation and good closed-loop performance over a wide range of delay variation. In a second experiment E M V R E is used to model a Reject Pulp Refiner. Although the measurements are very noisy, the delay estimation converges in less than 20 samples to the value obtained by offline analysis. The experiments demonstrate the applicability of the V R E delay estimation methods. 1.4 Thesis Outline C h a p t e r 2 shows the importance of correct delay estimation. The standard delay estimation methods are reviewed and the limitations of every method are highlighted. C h a p t e r 3 describes the idea of Variable Regression and how the E M V R E can be generated by modifying any standard parameter estimator. It is shown that under certain conditions, the delay estimation converges only to the correct value. C h a p t e r 4 introduces the F M V R E and shows how the model parameters are selected. Proofs of the estimation convergence, bounds, robustness and consistency in the presence of coloured noise are given. The excitation conditions required for accurate delay estimation are derived and it is shown that one sinusoid is sufficient excitation. Chapter 1. Introduction 6 C h a p t e r 5 presents the adaptive versions of pole placement, Dahlin and generalized predictive controllers designed based on the V R E . The analysis of delay estimation in closed-loop is also presented. C h a p t e r 6 shows the experimental results of applying the V R E to a laboratory temperature control experiment subject to a wide range of delay variation. Results from an identification experiment on an industrial reject pulp refiner are also presented. C h a p t e r 7 summarizes the results of the thesis, and gives suggestions for further research. Chapter 2 I n t r o d u c t i o n to D e l a y E s t i m a t i o n Many systems have a time-varying delay as well as time-varying dynamics. A good on-line estimation algorithm is required in these cases to track both the parameters and the delay changes. On the other hand, most control designs are sensitive to wrong delay assumption or estimation and can easily go unstable in such a case. This fact emphasizes the need for a fast and accurate on-line process delay estimation. One method to overcome the problem of unknown delay is to use an extended horizon controller or a generalized predictive controller designed for the maximum expected delay (e.g. Clarke et al. 1987). A major drawback of such a method is the unnecessarily detuned response that results. Consider the continuous system that has the transfer function: ^( ) _ nt.\ xr „-T G{s} _ = K..e' s 5 u { s n-l n > / « « + i ) t ¥=i r s d ) ' (2.1) n w ^ + i ) where: K„, is the steady state gain, Td is the delay, n is the number of poles, copi is the i'th pole corner frequency, and u) i is the i'th zero corner frequency. z The transfer function G(z) is obtained using the Z transform and a zero order-hold 7 Chapter 2. Introduction to Delay Estimation 8 as: G(z) = (i _ z 2 {^*} s - i ) (2.2) and has the form: Xlll ^^- \ ^\-, G U ) = where k = integeriTd/T,) i + i ^-" (2.3) is the delay in samples, and T, is the sampling time. Equation 2.3 can be written in the vector form: y(t + l) © = *(<)0, (2.4) = [ - y ( < ) , - - , - y ( < - n + l ) , u ( < - f e ) , . . - , u ( t - « ! - n + l)], (2.5) = • • • ,«n,&o, • • • ,b -i] T n (2.6) $(<) is the regression vector and 0 is the parameter vector. In this study only the elements in the parameter vector are called the "parameters" of the system. The system delay does not show up as an element in the parameter vector, but it shows up in the structure of the regression vector. Our objective is to estimate k with only prior knowledge about its bounds: and 0 can be either known or unknown. The following example demonstrates the importance of a correct delay estimation for control. Example 2.1 Consider the first-order system: 9 Chapter 2. Introduction to Delay Estimation which is estimated as: G(z) = z z — a and the required closed-loop transfer function is: G (z) r = z-' -^— z —a k The error feedback controller Gfb(z) required to obtain this closed-loop function is given by: and using this controller, the actual closed-loop transfer function is: G (z) cl GfbG l + GG fb The actual closed-loop transfer function (G i) equals the required one (G>) only if c we have a correct model: a = a, 6 = b and k = k. The root-locus technique can be used to study the effect of having errors in estimating a and b. It is well known that a small error in those parameters changes the location of the closed-loop poles but the poles can still be inside the unit circle. On the other hand, the root-locus technique cannot be used to study the closed-loop pole locations as a function of the delay estimate because the characteristic equation (1 + GfbG = 0) is not linear in k and because a change in k can also change the order of the characteristic equation. There is only a limited number of possible k defined by 2.7 and the roots of the characteristic equation can be calculated for all possible k. Assuming perfect knowledge of the system parameters d = q, b — b, the closed-loop transfer function is then: 10 Chapter 2. Introduction to Delay Estimation Consider for example that a = .36,/? = .64 and fc { = l , f c m n m a x = 5. The possible combinations of k and k give: k 1 2 3 4 5 1 - - * * * 2 - - - * * 3 - - - - * 4 * * - - - 5 * * * - - where "—" denotes a stable closed-loop system and "*" denotes an unstable or very oscillatory closed-loop system. Most of the stable combinations have highly oscillatory poles which give a poor closed-loop response. As the difference \k — k\ increases, the system can be easily driven to instability, even that we know the parameters a and b perfectly. One way to assure a stable system is to ask for a slower closed-loop response. If the required closed-loop system is twice as slow as the case above, we get a = .6,/3 = .4, and for these parameters we have a stable closed-loop for any combination of k and k. So, a detuned controller will be robust for errors in the estimated delay, but at the price of a slower closed-loop response. It is clear that a correct delay estimation is essential for a good closed-loop performance. 2.1 M e t h o d s of D e l a y E s t i m a t i o n In this section, the important delay estimation methods are reviewed. In on-line estimation, we have the extended numerator polynomial method and the rational approximation method. In off-line estimation we have the cross-correlation method. 11 Chapter 2. Introduction to Delay Estimation 2.1.1 T h e Extended Numerator Polynomial M e t h o d In this method, we replace the numerator polynomial in Equation 2.3 by an extended one and the expected output can be written in the form equivalent to that in Equation 2.4 as follows: y (t + l) e = $ (f) = e $ (<)0 , e e [-y(t),---,-y(t-n + l), u(t - k ), min • • • ,u(t - k max - n + 1)}, [a • • - , ( ! „ , 6 l5 fci The suffix e stands for the extended polynomial. There are additional (k max — k ) elements in the parameter vector. The first (k — k i ) elements and the last (k m n min — k) max elements in the parameter vector should be zeros in order to have y = y. e In order to estimate k and 0 , we have to estimate the extended vector 0 and e conclude the value of k from the first zero elements. The method is heavily used in practice (e.g. Biswas and Singh, 1978; Vogel and Edar, 1980; Kurz and Goedecke, 1981; Gerry et al., 1983; Chien et al., 1984; Chandra et al., 1985; Clough and Park, 1985; Lammers and Verbruggen, 1985 and in many others). The method has the following drawbacks: 1. This method essentially requires an unbiased estimate of 0 . This means that e the input signal should be rich enough to identify the 2ra system parameters plus the additional (k max — k i ) elements. m n 2. If there is coloured noise, then the estimated parameters will be biased and the delay cannot be concluded. This means that the noise model should also be estimated which means additional parameters to estimate and a slower convergence. Thus, the Recursive Least Squares (RLS) parameter estimator cannot 12 Chapter 2. Introduction to Delay Estimation be used and the Approximate Maximum Likelihood ( A M L ) estimator should be used instead. 3. The number of parameters in the extended numerator polynomial increases linearly with the system delay or the sampling rate. This means slower parameter convergence and longer time to identify the delay. 2.1.2 The Rational Approximation Method. In this method the term e~ TdS in Equation 2.1 is approximated by a low-order rational approximation Gd{s), that is: n>M; +1) G(s) ~ K„G (s)^ d Ift*M* +1) «=i The approximations include Pade approximation (Robinson and Soudack, 1970 and Gabay and Merhav, 1976), Walsh functions (Rao and Siuakumar 1979 and Rao 1983), Tchebycheff approximations (Knowles and Emre 1985), all pole approximation (Gawthrop and Nihtila, 1985), Laguerre series (Zervos, 1988a,b and Zervos et al. 1990) and all-pass niters (De Souza et al. in 1988). The approximation is selected such that it provides a good approximation of the system in the bandwidth of interest. The discrete time model of the system will not have an explicit value of the correct delay because the original system parameters and the delay approximation parameters cannot be decoupled. The rational approximation has the following drawbacks: 1. The structure of the rational function is not kept after sampling which means that the delay cannot be distinguished from the process dynamics. 2. Most of the approximations use non-minimum phase functions which restrict the control design choices to indirect methods that preserve the unstable zeros. 13 Chapter 2. Introduction to Delay Estimation 3- The poles of the delay approximation must be close to the system poles to insure that the sampling rate is suitable for identifying both the system poles and the delay approximation poles. This means that the order of the approximation should increase linearly with process delay. 2.1.3 T h e Cross-Correlation M e t h o d The main idea here is that the impulse response of the system is the same as the inputoutput cross-correlation if the input is white. The delay is then obtained as the time of first non-zero output on the system impulse response. The method is used mainly in signal analysis (Knapp and Carter, 1976; Hassab and Boucher,1979) or to estimate the bound on the delay variation [k i ,k ] m n max (Allison et al, 1989). The disadvantages of this method are: 1. The input must be white to have an impulse auto-correlation function. Any other input does not have a zero auto-correlation at all non-zero time shifts. The resulting cross-correlation function is then not zero for time shift less than the delay. A n d consequently the estimated delay will always be less than the true one. 2. The limit between what is considered zero and what is considered non-zero for the cross correlation function should be very carefully defined. This limit allows a non-white input to be used, but on the other hand this limit depends mainly on the system which is generally unknown. 3. The method is mainly for open-loop estimation because it does not force causality on the estimated system and the feedback path between the input and the output (the controller) may be estimated instead of the feedforward path (the system). Chapter 2. Introduction to Delay Estimation 14 4. The method is an off-line method because a large number of measurements should be collected before the cross-correlation function can be generated and the delay can be estimated. 2.2 Conclusions In this chapter, we demonstrated the importance of having an accurate delay estimation and showed that wrong delay estimation can easily destabilize the closed-loop. The standard methods of delay estimation are reviewed. The extended numerator polynomial, the rational approximation and the cross-correlation methods are discussed and their disadvantages are focused on. Chapter 3 Delay E s t i m a t i o n U s i n g V a r i a b l e Regression In this chapter, we present a new method for the simultaneous recursive estimation of both process parameters and delay. The new algorithm is a modification of any standard recursive parameter estimation algorithm and thus is easily applied to any estimation algorithm such as Least Squares, Instrumental Variables, Maximum Likelihood,...etc. and their extensions. The new algorithm requires minimum additional computations compared to the original algorithm. The delay is explicitly estimated and the parameter vector has the same length as for the known delay system and no additional excitation is required. This means clear separation between the system dynamics and the delay. 3.1 V a r i a b l e Regression The basic idea in variable regression delay estimation is to use a first-order model to estimate the delay. The estimated delay is then used to modify the structure of the regression vector and a variable regression is generated. Let the system be described by the A R M A X equation: y(t) = ${t,k)e + w(t) (3.9) where: 0 H / ( * - 1 ) , — y(t — n),u(t — k — 1 ) , u ( t — k — n)} (3.10) [a-i,a , (3.11) n &o, •••,&n-i] T 15 16 Chapter 3. Delay Estimation Using Variable Regression k: system time delay n: number of poles w(t): coloured noise. The parameter vector (0) is a function of the parameters, while the measurement vector ($) is a function of the time delay. Since many industrial processes are overdamped and can well be approximated as first-order plus delay, the model used is first-order plus delay. It will be shown that the first-order model is enough to estimate the delay even if the system dynamics are represented by a higher-order model. The estimated delay can then be used in the estimator of the high-order model parameters. Let the model parameters and the estimated time delay be 0 m and k respectively. The predicted output is given by: y = $(«,fc)0 (3.12) = [a ,b ] (3.13) ; = [- (t-l),«(«-fc-l)] (3.14) m where: 0m where 0 < — a m T m m y < 1. k can take any integer value in the horizon [fc j , k ]. m n max The structure of the regression vector $ is variable depending on the estimated value k. This variation of the regression vector characterizes this method and hence the designation is "Variable Regression Estimator (VRE)". The prediction error is defined as: e(t, 0 , k) = y(t) - y{t) = y{t) - k)Q m m (3.15) The performance index is chosen to be: j(t, e , k) = - jy{i, m * i=0 0 , k)\ m 2 (3.i6) Chapter 3. Delay Estimation 17 Using Variable Regression The performance index is a function of both the model parameters and delay. So in order to minimize the performance index, Equation 3.16 should be minimized with respect to both the model parameters and time delay. The solution will identify 0 m and k. This means that: (3.17) 'max (3.18) Equation 3.18 is not written in the partial derivative form because k takes only discrete values. This characteristic of the delay equation makes the solution of 3.18 robust to modelling error in 0 m and/or the existence of noise (white or coloured). If the model has the correct system parameters and in the absence of noise, Equation 3.18 will have a minimum of zero. If there is a small modelling error and/or process noise, Equation 3.18 alone still can be enough to identify the delay correctly. The following Theorem shows that in the case of known parameter system, the delay can be accurately estimated using Equation 3.18 alone. T h e o r e m 3.1 ( K n o w n S y s t e m parameters) For any stable system of known parameters ( 0 m = 0 ) , if the measurement noise is uncorrelated with the system input, then J(k) is minimum when k — k. Proof: Let: y(t) $(t,k)e + w(t) $(i,fc)0 then: e(t, k) = ($(*, k) - fc))0 + w(t) Chapter 3. Delay Estimation Using Variable Regression 18 The performance index is then: J&k) = 7EKO] 2 = 7 £> (01 + K*fr ) 2 + [ M O W , *) - fc fc))e] As the number of measurements increases, the different terms in the last equation can be replaced by their expected values, then: J(oo,fc) = lim J(t,k) The term E(w (i)) is by definition equal to T*u,(0), the auto-correlation function of 2 w at zero shift. The last term should be zero because it is a summation of terms of the form E(w(i)u(i — j)) and this is by definition equal to r {j), the cross-correlation uw function of w and u at shift j and this cross-correlation is zero at all shifts because the two signals are uncorrelated. The second term is the squared summation of the input auto-correlation function values at different shifts and this squared summation cannot be negative. From the above analysis, we conclude that J(k) = r {0) for k = k, and J(k) > r (Q) w w for k ^ k. There are only two cases for which J(k) — r (0) w and k ^ k. The first case is when the input has a flat auto-correlation function {r (T) = r (0)Vr). This corresponds u to a constant level input (u(t) = constant, u k) = The system will have a steady output and will look the same for any value of the delay. Any input change is sufficient for correct delay estimation. However, a richer input will have a sharper auto-correlation function and a sharper minimum of J(k) at k = k. The second case for J(k) = r (0) w and k ^ k is when the input signal is periodic. In this case J(k) is also periodic and have the same period T , i.e. J(k + T /T„) e e = J(k). The input signal 19 Chapter 3. Delay Estimation Using Variable Regression should be so slow that J(k) has only one minimum in the possible range [^ i ,fc ]ra T must then satisfy that T /T„ > (k e e max n mai — fc ). This condition defines the lower limit mtn of the excitation period. This condition is very general and not restricted to systems of known parameters. • The result of this theorem is valid for any system order, provided that the model has the same order and parameters as the system. Any non-constant input is sufficient and for periodic input the condition given above should be satisfied. The case of unknown parameters is studied in the following theorem for the first-order systems. T h e o r e m 3.2 ( U n k n o w n S y s t e m Parameters) / / the input is white and uncorrected with the coloured measurement noise, then minimizing J gives an unbiased delay estimation for a first-order stable system of unknown parameters. Proof: Let: y(t) •= -ay(t - 1) + bu(t - 1 - k) + w(t) y(t) = -a y{t - 1) + b u(t - 1 - k) m m then, e(t, k) = (a - a)y(t - 1) + bu(t - 1 - k) - b u(t - 1 - k) + w(t) m m A similar analysis to that in the known parameters case gives: J(oo, 0 , k) m = [(a - a) r {0) + (b + b )r (Q) + r (0) + 2(a - a){br (k) + r (l))] 2 m -2 b 2 y m 2 m u w m uy wy [(a - a)r (k) + br (k - k)] m uy u The terms in the first square brace do not depend on k, and so, only the two terms in the second square brace should be considered in the minimization. For a white 20 Chapter 3. Delay Estimation Using Variable Regression excitation: 0 r (0) i f r = 0 u ru(T) = \ },r {T) 0 and since \a m 6r (0)|a|- if r > 0 — o| < 1, then ( a m iir <k ) = uy T+1+fc u — a)r (k) uy if r > fc J < br (0) and J(k) is minimum at k = k, u and this proves the theorem. • It is to be noted that having a white input is not a necessary condition, but rather a sufficient condition. As the difference \a — a\ decreases, the minimum of J(k) occurs m at k = k for other inputs not as rich as the white input. It must be noted that we assume a positive value for a, which is the case for any physical system. E x a m p l e 3.1 Consider the first-order system: y{t) = -ay{t - 1) + q~ bu(t - 1) + w(t) k where a = —.5,b = .5 and k = 3 and u is a step input. Three tests are carried out and the results are summarized in Figure 3.1. The three cases studied are: 1. Correct parameter estimates ( 0 has a sharp minimum of zero. m = 0) and no noise [w = 0). Equation 3.18 Figure 3.1 shows a family of curves, each of them represents a specific time instant with data collected only up to that point. Curves representing time intervals less than the delay will have a flat minimum and reflect the poverty of the information collected. 2. 50% error in the estimated parameters (d = —.25, b = .75) and no noise. Equation 3.18 alone is still able to identify the delay accurately, although the input is not white. The minimum is not zero any more and this usually indicates the inaccuracy in the estimated parameters. Figure 3.1: Minimization of J 22 Chapter 3. Delay Estimation Using Variable Regression 3. As in the case above but with additional measurement noise (w(t) = N(0, .1)). Even under this condition, Equation 3.18 alone is still enough to identify the delay accurately. The last two theorems imply that there is a strong correlation between the excitation requirement of the input signal and the uncertainties of the parameters. The more accurate the parameter estimates are, the less is the excitation required for correct delay estimation. Since the system parameters are generally unknown, we have either one of two choices: 1. Estimating the model parameters that best describe the system. In this case Equations 3.17 and 3.18 should be solved simultaneously. This case will be called the Estimated Model Variable Regression Estimation ( E M V R E ) . Both the model parameters and the delay are to be estimated recursively. 2. Using fixed model parameters to estimate the delay by Equations 3.18 only. The fixed model should be chosen such that it is most likely to have correct delay estimation. This method will be called the Fixed Model Variable Regression Estimation ( F M V R E ) . The delay is estimated recursively and the estimated value is fed to the variable regression vector of another model used to describe the system and whose parameters are to be estimated recursively. 3.2 T h e E s t i m a t e d M o d e l V a r i a b l e Regression E s t i m a t i o n ( E M V R E ) In this case both the model parameters and the delay are estimated on-line. Since good estimation of the model parameters results in a good delay estimation, it is logical to use the most recent parameter estimates to estimate the delay. Similarly, since good parameter estimation requires good delay estimation, it is logical to use the most Chapter 3. Delay Estimation Using Variable Regression 23 recent delay estimate to estimate the parameters. The implementation of the model parameters and the delay estimation is carried out in two steps. In the first step the standard Recursive Least-Squares estimator is used to estimate the parameters using the last estimated delay. The next step is to estimate the delay using the last estimated parameters. The two steps are to be carried out once every sample. The steps of the modified version of the standard Recursive Least-Squares estimator will be: 1. Start with initial guess: 0 m = 0o and k i m n < k< k . max 2. Generate the corresponding regression vector 3. Update the parameter estimation ( 0 ) using RLS. m 4. Update the performance index values (J) over the horizon [k i , k ] m n max using the recent parameter estimates. 5. Estimate the delay k that minimizes the performance index. 6. New sample: go to 2 above. The equations required to implement those steps are as follows: K(t + 1) = P(t)$ {t + 1, k(t))/[\ + ${t + 1, k{t))P(t)$ {t + l,k(t))] T (3.19) T e (t +1) = e (t) + K{t + i)( {t +1) - $(< + 1 , k{t))e {t)) m m y P(t + l) = [I-K(t (3.20) m + l)${t + l,k{t))]P{t)/\ J(t + 1,1*) = XJ(t, 1*) + [y(t + 1) - $(* + 1, fc )0 (i)] VA 2 i m ;i (3.21) € [k , k ] min J(t + l,k(t + 1)) = minj(t + 1, k^ki € [k , k ] min max max (3.22) (3.23) where P is the parameter covariance matrix, K is the gain matrix, and A is the forgetting factor. Chapter 3. Delay Estimation Using Variable Regression 24 Equations 3.19-3.21 are the standard Least-Squares estimator equations for a constant k. The additional two equations are the delay estimation equations. The implementation of the additional equations requires minimum storage and computation time as they contain only simple multiplications and additions and a simple search routine. Equations 3.22 and 3.23 can be simply added to any recursive estimator. Computer simulations show that the delay estimation does not have to change with every change in the parameter estimation. In other words, the delay estimation will not change at every sample as the parameters will. As more data are collected, the parameter changes decrease towards the final convergence values and at the same time the period at which the delay estimation is not changing increases. As the parameters converge to constant values, the delay estimation will reach its correct value and leave the parameter estimation to converge to the correct values uninterruptedly. This characteristic suggests resetting the parameter covariance matrix (P) every time the delay estimation changes to speed up the parameter estimation convergence. If the delay estimation converges, the parameter estimation will have the same steady-state properties as the original estimator. It is difficult to derive the general convergence properties of this method because of two reasons. The first is the mutual interaction between the delay and the parameter estimation which means that the delay and parameter estimation equations should be solved simultaneously. The second reason is that the delay updating equation is not in a difference equation form as in the case of the parameter estimation equations. However, for a first-order system we can prove that if the algorithm converges, it will converge only to the correct delay. In other words there is only one convergence point for the algorithm. T h e o r e m 3.3 (Convergence of the E M V R E ) cited by a white input, if the EMVRE For a first-order stable system ex- delay estimation converges, it converges only to 25 Chapter 3. Delay Estimation Using Variable Regression the correct value . Proof: Let the algorithm converge to the values a ,b ,k, m then from Theorem 3.2 we m have k — k provided that the excitation signal is rich enough. • The estimated parameters are essentially those obtained by R L S using the correct delay. These parameter estimates can be biased in the presence of coloured noise. However the Approximate Maximum Likelihood can be modified with the variable regression technique to guarantee unbiased parameter estimates. Although the result of this theorem is restricted to white input, computer simulation and real-time experiments show that correct delay estimation can be obtained with a much smaller level of excitation. The following examples show the characteristics of the E M V R E algorithm. In the first example, a first-order deterministic model is considered. The parameters and the delay have large simultaneous changes. The second example shows the E M V R E working in an adaptive pole placement loop. The third example shows the estimator behaviour over a wide range of sampling rate variations. In the last example, the E M V R E is used as a method of model reduction. E x a m p l e 3.2 (Variable P a r a m e t e r s a n d Delay) Consider the system: y(t) = -ay{t - 1) + q- bu{t - 1) k The input u(t) is a square wave with period of 40 sample, a, 6, k change with time as follows: Time a b k 0 < t < 200 -.50 .50 3 200 < t < 400 -.25 .75 1 400 < t < 600 -.75 .25 5 Chapter 3. Delay Estimation Using Variable Regression 26 The simultaneous large changes in the parameters and delay represent a good evaluation test for the proposed estimation method. The results in Figure 3.2 show a fast and accurate tracking of the changing parameters and delay. E x a m p l e 3.3 ( A d a p t i v e pole placement) Many adaptive controllers have been simulated successfully using the E M V R E . Some adaptive controllers are very sensitive to correct delay assumption or estimation. The pole-placement controller is one of those delay-sensitive controllers. The system used in Example 3.2 is now reused with an adaptive pole placement controller with a closedloop time constant of two samples. The controller design can be found in Chapter 5 where detailed analysis of the adaptive controller design is given. The simultaneous large changes in the parameters and delay would drive the system to instability if a non-adaptive controller is used, the results of the adaptive controller based on E M V R E are shown in Figure 3.3. E x a m p l e 3.4 (Variable S a m p l i n g T i m e ) Consider the system: G(s) Where Tj, — lsec.,r = lsec -TiS TS 1 + The sampled system has the form of Example 3.2. The sampling rate directly affects the values of the parameters and the delay. Three sampling rates are used to test the estimator. The model equation has the following parameters as the sampling rate (JP„) changes: F, a b k 1 Hz -.3678 .6321 1 5 Hz -.8187 .1813 5 10 Hz -.9048 .0951 10 Chapter 3. Delay Estimation Using Variable Regression O. 5 it 27 4.. •yA-iUr. Tip w r *v 1 • i i -O. 5 -1 .<? 0.0 200. Add, •2. 0.0 600. 0. 8 2<?0. 400. Q<?<? 400. 6QQ r— 0. 6 I I I -0.5 < . O D. 4 0. 2r -1 .-2 200. 4(?<J. 600. 2 0. c- SAMPLES Figure 3.2: E M V R E of a first order system. 200. SAMPLES Chapter 3. Delay Estimation Using Variable Regression Y 0. - l. 0.0 . O.OQ 200. a. ] 600. Add. 8i 0. 0 I. Q 200. 400. - b 1 1 q -0.75 •L.5<? Q.Q 200. 4<?<?. SAMPLES 600. 0.0 200. 400 SAMPLES Figure 3.3: Estimation and adaptive P-P control using the E M V R E . Chapter 3. Delay Estimation Using Variable Regression 29 The results in Figure 3.4 show excellent estimation of the true values of the parameters and of the time delay for the different sampling rates. The input signal used was a square wave. E x a m p l e 3.5 ( H i g h order System) The system has a fourth-order transfer function: G(s) s + 2 W 2 Vl lS + W? s + 2v W s + W Where: 77! = 0.5, W^i = l.Orad/sec,rj 2 2 2 2 2 2 = 0.75,^2 = 2.0rad/sec. the step response of the fourth-order model and the second order dominant model are shown in Figure 3.5 as Y4 and Y2 respectively. It is clear that the fourth-order model response can be represented by a delayed second order one. The sampling rate is 10Hz and a second order system with delay is estimated using a square wave input. The step response of the estimated model is shown in Figure 3.5 as Y. This simulation shows that the dominant poles are not enough to describe the system and a delay is a good approximation of the unmodelled dynamics. This process is the opposite of approximating the delay by a rational function and is a common industrial practice. 3.2.1 Advantages and disadvantages of the EMVRE The E M V R E has unique advantages over both the extended numerator polynomial and the rational approximation. The main advantages of the E M V R E are: 1. Since no parameters are added to the process dynamics model, the parameter convergence is fast and no additional richness conditions are required from the input signal to estimate the model parameters. 2. Unbiased delay estimation can be obtained in presence of coloured noise even with biased model parameters. Chapter 3. Delay Estimation Using Variable Regression 30 5 HZ 1 HL 10 HZ o. •1. 1.0 50. 100. 0.0 IO0. 200. 50. 100. 07 0 IO0. 200 0.0 50. 100. 0.0 IO0. 200. -. i 0.0 0 100. 200 0.5 I i O.Q -0.5 -t .0 0.0 100. 200 I .0 i l < JO 0.5 O.O -0.5 0.0 SAMPLES SAMPLES Figure 3.4: E M V R E at different sampling rates. LOO. 200. SAMPLES Chapter 3. Delay Estimation Using Variable Regression l .5 —Y2 --Y4 I .<? 31 — U l. 0 0. 5 I 0.5 o.o -0.5 0.0 1.5 -0. 5 50. 100. 150. 200. -1.0 0.0 200. 400. 600. 800 0.0 200. 400. 600. 800 Y 4 •Y I .0 O. 5 O.O 0.5 0.0 50. 100. 150. 200. SAMPLES Figure 3.5: E M V R E of a high order system. SAMPLES Chapter 3. Delay Estimation Using Variable Regression 32 3. The additional equations required to estimate the delay are easy to implement with minimum code, computation time and storage. The E M V R E main disadvantages are: 1. There is no proof that the algorithm will always converge. 2. There is no general result on the method behaviour with system order higher than one. 3. The exact excitation requirement for correct delay estimation is not known. The disadvantages of the E M V R E could be overcome with the introduction of the F M V R E in the next chapter. 3.3 Conclusions In this chapter, the new idea of variable regression is presented and used to develop the Estimated Model V R E . The V R E is combined with Least-Squares parameter estimator to form recursive estimation of both the system parameters and delay. For systems with known parameters, the delay is estimated with the minimum possible excitation. For first-order systems, the E M V R E can converge only to the true values when rich excitation is used. Chapter 4 The F i x e d M o d e l Variable Regression E s t i m a t i o n ( F M V R E ) The F M V R E uses a fixed parameter model to estimate the delay. It will be shown that the best results are obtained when the first-order model pole lies on the unit circle (an integrator). The method in this special case is shown to be equivalent to maximizing the cross-correlation between the input and the output increments. This is compared with the standard cross-correlation method and the advantages are clarified. Consider the stable system: U(z) 1 + z-* + a z~ + • • • + a z2 W ai n 2 V ; n where: k > 0 is the system time delay, and n is the system order. G(z) can be expanded as follows: G(z) = z- [-^— + + ••• + -5—] z - Pi z- p . z-p h 2 n where c; and pi are real or complex numbers and \pi\ < 1 V i € [1, n], G(z) can be expanded as an infinite series: G{z) = z- - [ h (l+p z- +plz- +plz- 1 1 C l 2 + ---) 3 1 + C (1 +P2Z- +c (1 + p -i 2 n 1 nZ 33 +p Z~ 2 +p Z~ 2 3 2 3 2 + •••) + pi*- + plz-z + ...) ] 2 (4.25) Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 34 The impulse response coefficients of this system are: 0 if (t - 1) < k JT« ^impulseW = ~< if (t-l) ±c r { iP yk =k (4.26) if ( < - i ) > fc i=l Using a first-order plus delay model, the estimated output is defined as: y{t) = q^a^t) and 0 < a m + q-' - b u{t) k (4.27) l m < 1. The prediction error is given by: e 2 (4.28) y(t)-y(t)) =( 2 = ( y(t) - a y{t - 1) - b u{t m -l~k)f m (4.29) = ( y\t) + a y (t - 1) + b u (t - 1 - k) - 2a y(t)y(t - 1) 2 2 2 m 2 m m -2b {y{t)u(t - 1 - k ) - a y{t - l)u(t - 1 - k) ) m (4.30) m The performance index J ( 0 ,fc)is equivalent to the expected value of e and is given 2 m by: J{e ,k) m = E{c ) = E -2b E 2 0 m (4.31) 1 where: E 0 = (l + a Jr (0) Ei = r ( l -ffc)- a r (fc) and r (0) — E(y (t)),r (k) 2 y uy 2 y u y + m b r (0)-2a r (l) (4.32) 2 m u m y (4.33) uy = E(y(t)u(t — k)). The value of E does not depend on 0 k, so in order to minimize J(Q ,k) m we have to maximize E\ with respect to k. In the analysis given here we assume, without loss of generality, a positive static gain system. Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 35 For systems having negative static gains, E\ should be minimized. From the impulse response coefficients of the system, we conclude that for white excitation: if k < k 0 if k = k r (l + k) = r (0) < u uy i=l n Y^ iPi~ c if k> k k and r {k) = r (0) < uy u 0 if k < k 0 if k = k fc-fc-i (4.34) if k > k then: if fc < k 0 n E x if fc = Jfc = r (0) < u i=i n X)cip* " f c - a c p^- k m ii 1 i (4.35) k>k i=l The terms in the last equation are very similar to the terms of the impulse response coefficients of the system. In fact they are identical to the impulse response coefficients of the function: Gi(z) = (1 - U(z) (4.36) a z- )G{z) x m which has the impulse response coefficients: V\ impulse^) = 0 if (t - 1) < k I > if (t-l) = k i=l ±c I i i=l P t 1 ] - k - a c^'- - 1 ) m f c 1 if (*-!)>* (4.37) Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 36 The impulse response of the function Gi(z) is identical to the response of the original system G(z) when the input is: U {z) = 1 - a z x 1 Ul (4.38) x m (t) -a m 0 if 0 < t < 1 if 1 < t < 2 (4.39) if 2 < t The last conclusion is important because it means that the value of k that maximizes Ei can be determined from the system response to the input sequence defined in the last equation. This is illustrated in Figure 4.6 in which a general overdamped system having k — 4 is considered. The response of the system is shown and related to the function Ei(k). In Figure 4.6 the maximum value of E\ occurs at k ^ k. That is because the estimated value of the delay depends on the values of a m and the sampling time. From Equation 4.35, there are two conditions for unbiased delay estimation (i.e. E\ is maximum at k = k): C o n d i t i o n 4.1 X>>o This condition means that the first non-zero value of the system impulse response is of the same sign as the steady state gain. This means that regardless of whether the system is minimum or non-minimum phase, the sampling rate is slow enough to have the first non-zero sample positive. C o n d i t i o n 4.2 n ]T)ci i=l n > J^Cipl - a C i p - , V n > j > 0 _ 1 m i=\ Chapter 4. The Fixed Model Variable Regression Estimation Figure 4.6: Impulse response of G x (FMVRE) 37 Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 38 This condition means that the first nonzero value of the system response to the input given by Equation 4.39 is the maximum of all the samples. influenced by the value of a m The response will be that determines the amplitude of the negative input in Equation 4.39. The case when a m = 0 is the most difficult one and is equivalent to maximizing the cross-correlation between the system input and output. In this case the condition for unbiased delay estimate is that the sampling rate is slow enough to have the first nonzero sample of the impulse response as the maximum value of all the samples. This is always true for first-order systems. 4.1 T h e F i x e d M o d e l Parameters The other extreme case is when a m = 1. In this case the input in the second sampling period is at its maximum negative value. This means that the second non-zero sample will have an amplitude much less than in'the C3LSC O/jyi = 0 and consequently it is most likely that the first non-zero sample will be the maximum. The method is equivalent to maximizing the cross-correlation function between the system input and the output increments. The value of b is no longer important because we are concerned only with m the maximization of E whichr now takes form: {l + k)- the r (k) r uy uy E((y(t)-y(t-l))u(t-l-k)) E{Ay(t)u(t (4.40) -l-k)) where: A = (l Step disturbances and/or changes in the set point are automatically eliminated in Ay which means that they do not affect the delay estimation. From Equations 4.26 and Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 39 4.37, E i can be expressed as: Ei{k + 1) yumpuheit) = = Ayimpulse{t) = AAy {t) = (4.41) = A y.te {t) 2 Hep P (l-2q- +q- )y (t) 1 2 ttep The advantage of relating the function E to the step response is that the step response x does not change with changing the sampling time, while the impulse response does change. This leads us to the bounds of the delay estimation: T h e o r e m 4.1 The time delay estimated by the FMVRE is bounded for any sampling rate. Proof: For very large sampling time, the sampled step response is given by: y.te (0) = o P (* ^ n\ y.te {t > 0) P fro = y„ = V frn-l + &1 H —— 1+ a + • • •+ a x n Then: 0 if k + 1 = 0 y„ if k + 1 = 1 -y„ if k + 1 = 2 0 if k + 1 > 2 Ei Then the maximum of E occurs when k = 0, and this defines the lower bound of the x estimated time delay. On the other hand, for very small sampling time, the operator A approximates the time derivative. The operator A 2 approximates the second time derivative. This means that E in Equation 4.41 is equivalent to the curvature of the x step response curve: E i ^ T ^2 2 y (t) etep Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 40 Then the maximum of Ei will occur at the point of maximum curvature of the step response and this defines the upper bound of the estimated delay. For any value of T s such that: 0 < T < oo, the operator A is finite and so is the value of Ei, and this a proves the theorem. • The upper bound on the delay estimate is closely related to the delay approximation given by the Ziegler-Nichols method (Astrom 1988). In the Ziegler-Nichols approximation, the time delay is defined as the intercept of the maximum slope line with the time axis. For minimum phase systems, this time is usually very close to the time of maximum curvature as shown in Figure 4.7. The importance of this relation is that controllers tuned based on the Ziegler-Nichols approximation can now be tuned on-line to form an adaptive closed-loop. 4.2 E s t i m a t i o n of the S y s t e m P a r a m e t e r s The delay estimated using the F M V R E is independent of the other parameters of the system which are generally unknown. To estimate the parameters of the system, we need another model whose parameters are in 0 . Standard RLS can be used to estimate 0 with a variable regressor $(&). The estimation of the delay does not depend in any way on the vector 0 , and the resulting decoupling between the delay and the parameter estimation gives the F M V R E several advantages over the E M V R E . The modification of the R L S is given here and similar modification can be applied to other estimators. 1. Start with initial guess: 0 = 0 and k O min <k< k . 2. Generate the corresponding regression vector $(&). 3. Update the parameter estimation (0) using RLS. 4. Update Ei over the horizon [k i ,k \. m n max max Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) Figure 4.7: Delay estimation using Ziegler-Nichols and F M V R E . 41 Chapter 4. The Fixed Model Variable Regression Estimation 42 (FMVRE) 5. Look for.fc corresponding to the maximum of E\. 6. New sample: go to 2 The equations required to implement the R L S / F M V R E are: K{t + 1) = P{t)$ {t + 1, fc(t))/[A + $(< + 1, k{t))P{t)$ (t + 1, k{t))] T T e(t + i) = e{t) + K{t + i)(y(t + i ) - ^ t + i,k{t))e{t)) P(* + 1) = [/ - + + 1, fc(i))]P(t)/A E [t + l,ki) = XE^t,ki) + u(t - ki)[y(t + 1) - y(t)]Vki € [fc x £ ( i + 1, k(t + 1)) = moajjB^t + 1, ki)\/ki G a min ,k max ] (4.42) (4.43) (4.44) (4.45) [fcmin, (4.46) where P is the parameter covariance matrix, K is the gain matrix, and A is the forgetting factor. The properties of the modified estimator are essentially the same as those of the standard estimator. The reason is that the delay estimation converges independently of the parameter estimation. As the delay estimation converges to its final value, the regressor is no longer variable and the estimator returns to its standard version. Important special cases of delay estimation result from the above general discussions. 4.3 F i r s t - O r d e r Systems Theorem 3.2 proves that for white excitation, the correct delay estimation is guaranteed. This result is of practical importance because many industrial processes behave as first-order systems with delay and for these systems unbiased delay estimation can be guaranteed. As stated earlier, having a white input is not a necessary condition but rather a sufficient condition. It will be proven later that for the F M V R E , one sinusoid can is a sufficient excitation for delay estimation. The following example demonstrates the F M V R E method with a first-order system of variable parameters and delay. Chapter 4. The Fixed Model Variable Regression Estimation 43 (FMVRE) E x a m p l e 4.1 Consider the system: -T S_ G(s) = d TS + 1 The system time delay and time constant (in seconds) change with time as follows: Time T d T 0 < t < 100 7 2 100 < t < 200 4 5 200 < t < 300 1 8 The input u(t) is a square wave with period of 20 samples and T, —' lsec. Figure 4.8 shows the estimation results in open-loop and in closed-loop using a pole placement controller with a closed-loop time constant of lsec. The controller design is given by Equation 2.8. The delay estimation converges to the true values in less than 20 samples. 4.4 S e c o n d - O r d e r Systems It is well known (Astrom 1984-a) that for good parameter estimation and closed-loop control, the sampling time (T„) is related to the rise time (T ) by: r T / 2 > T. > T / 4 r (4.47) r In this section we show that the F M V R E does not restrict the sampling rate choice for second-order systems with reasonable damping. Consider the general second-order system: G(s) bs + 1 s /wl + 2(s/w + 1 2 0 Chapter 4. The Fixed Model Variable Regression Estimation 44 (FMVRE) 1 ST ORDER T P p 1 i _ ji 1 P ii ii i, 1 ii I i L "tJ ii ii , p L h o L _ Q.Q 20. , > i F H R — 1 — — f — 1 >i ii ii ii 1 ' i i 10. 60. 80. 100. 120. 140. 160. 180. 200. 220. 240. 260. 280. 3fl<?. OPEN-LOOP ESTIMATION h.Q 20. 4/>. 60. SO. 100. 120. 140. 160. 180. 200. 220. 24<?. 260. 280. 300. CLOSED-LOOP • 1 1 1 / / J 1 I I 1 1 1 ESTIMATION 1 I 1 I I 1 1 h i n n a n n n. h n n n n , u u U v/ \/ U U U li u u u u ai I 1 i i i i \ i 140. i i i i i i i i •3.0 20. +& . 60. 30. 100. 120^ 160. ISO. 200. 220. 240. 260. 28.?. 3-G'O. SEC . Figure 4.8: F M V R E of a first order system Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 45 where ( is the clamping ratio and for simplicity assume that the system time delay is an integer multiple of the sampling time. The minimum sampling time required for unbiased delay estimation is shown in Figure 4.9 as a function of ( for different values of b and w — 1. Given b and £, the minimum sampling time curve was obtained numeri0 cally by studying the system step response at different sampling rates and looking for the minimum sampling time that satisfies Condition 4.2. It is noted that for b > 0, the minimum sampling time required becomes very small and approaches zero as b becomes > .1, as shown for the cases b = 0.5 and 6 = 1. Figure 4.9 shows that for overdamped (( > 1), minimum-phase (b > 0) system, unbiased delay estimation is guaranteed if the sampling time satisfies Equation 4.47. The condition given on the sampling rate is not necessary but rather a sufficient one, as can be easily seen in Figure 4.9. Figure 4.9 shows that unbiased delay estimation is also achievable for under-damped systems. If the system has a minimum-phase zero (6 > 0), then any sampling rate is adequate for unbiased delay estimation regardless of the system damping ratio. If the system has no zero (b = 0), then the sampling rate required for unbiased delay estimation is still in the range defined by Equation 4.47 for £ > .5. Systems having £ < .5 are not very common in industry and for these systems the sampling rate requirements imposed by the F M V R E is usually slower than that required for estimating the other parameters. E x a m p l e 4.2 Figure 4.10 shows the delay estimation for the following parameter variation: Time T C b w 0 < t < 100 7 1.0 0 1 100 < t < 200 4 1.5 1 1 200 < t < 300 1 1.0 1 1 d 0 Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 6 = 0 b = 0.5 6= 1 Figure 4.9: Minimum T required for a second order system. t 46 Chapter 4. The Fixed Model Variable Regression Estimation 47 (FMVRE) The sampling time is one second which satisfies Equation 4.47 for the range of parameter variations. Figure 4.10 also shows that the condition of white excitation is a sufficient one because the delay is correctly estimated using a square wave input. The controller used is the no zero cancellation pole-placement given in Chapter 6. The delay estimation converges in less than 20 samples. 4.5 H i g h e r - O r d e r Systems Systems of order higher than two have no general form to study and no exact condition for unbiased delay estimation is achievable. However, high-order systems in industry are often over damped. In practice, these high-order over-damped systems are approximated by a low-order-plus-delay model. Our goal now is not to have an unbiased delay estimate of the system but rather to have a biased delay estimate that well approximates the high-order dynamics and the F M V R E automatically gives this approximation. Consider the system: ec) = <-h)i • <- > 4 48 which has no time delay (k = 0). The system has the step response shown in Figure 4.11. The sampling time is selected to be T, = lsec. For this sampling time, Condition 4.2 is not satisfied because , as shown in Figure 4.11, the maximum of the function Ei occurs at the fifth sample (k — 4) instead of the first sample. However, using the Ziegler-Nichols method we can obtain a good system approximation as a low-order system plus four seconds time delay (k = 4). This is exactly the value obtained from the F M V R E method. On the other hand, for white input, the standard cross-correlation between the input and the output will have the relation shown in Figure 4.11 which is also the impulse Chapter 4. The Fixed Model Variable Regression Estimation , O.O 1 p 1 2 ND ORDER 1 1 1 1 1 1 48 1 1 T 1 n n u ~1 1 i (1.0 1 (FMVRE) 20. ID. i 60. 80. r 100. 120- 140. 160. 180. 200. 220. 240. 260. 280. 300 1 OPEN-LOOP 10. -i i 1 i ESTIMATION r n i. _i i i i i i i _ i i i n i n n n i _ 0.0 20. 40. 60. 80. 100. 120. 140. 160. 180. 200. 220. 240. 260. 280. 300 CLOSED - LOOP ESTIMATION 10. ft. ~i r i! .1 \,A\ V" _J I SEC. Figure 4.10: F M V R E of a second order system. I L_ Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) Figure 4.11: Response of a high order system. 49 Chapter 4. The Fixed Model Variable Regression Estimation 50 (FMVRE) response of the system. The system delay equals the time to the first non-zero sample of the impulse response. In practice, the limit between what is considered zero and what is considered non-zero highly depends on the particular system under study. In our case, the limit is over 50% of the maximum value in order to obtain the same good approximation as by maximizing E\. It is clear that looking for the maximum of the function E\ is more practical than looking for the first non-zero of the cross-correlations function. The simulation results for this system are shown in Figure 4.12. The delay estimation converges in less than 30 samples. All of the preceding analysis was for noise-free systems. However, in open loop, all the results and conclusions are still valid even in the presence of coloured random noise as shown in the following analysis. 4.6 Effect of R a n d o m Noise T h e o r e m 4.2 The estimated delay (k) that maximizes the function E\ is unbiased in presence of process noise that is uncorrelated with the system input. Proof: Assume that the Gausian random white noise e(t) = N(0,a) affects the mea- surements through the stable filter G , i.e.: e y(t) where y u = G(q-*)u(t) = yu + G.{q-*)e{i) + (4.49) y e is the output due to the input and y is the output due to the noise. The e input-output cross-correlation is given by: r (t)=r (t) uy The term r u y u uyu + r it) (4.50) uye represents the deterministic part of the cross-correlation function and is given by Equation 4.34. To solve for the stochastic part we first expand G^q" ) as an 1 Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 51 8 TH ORDER >- 20. 4-fl. 60. 80. too. 120- 140. 1 6 0 . ISO. C. \1 M OPEN-LOOP ESTIMATION 2-20 CLOSED - LOOP ESTIMATION 0. •5. SEC. Figure 4.12: F M V R E of a high-order system. 2 00 Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 52 infinite moving average polynomial similar to that of G(z): G (q 1 e ) = w + w g +w q 1 0 1 2 2 (4.51) + --- that means: (4.52) y (t) = w e(t) + e(t - 1) + w e{t - 2) + • • • e 0 Wl 2 then the stochastic part of the cross-correlation function is given by: r. (0 w r (t) 0 ue 0 + tuir (< + 1) + w r (t + 2) -| ue + 2 0 + ue 0 (4.53) + ••• since e and u are uncorrelated. This means that the noise has no effect on the crosscorrelation function and consequently it has no effect on the function E\ and this proves the theorem. • This result means that the presence of coloured noise does not affect E\ in any way, even by increasing or decreasing E\. In the E M V R E where we minimize J , we showed that the coloured noise will increase J and so for small signal to noise ratio the maximum of J will not be sharp. While in the F M V R E method, the signal to noise ratio does not affect E\ and it is only the excitation richness that determines E\ for a specific system. The result of this theorem is very important because it guarantees unbiased delay estimation under the same conditions that will surely result in a biased delay estimation using least-squares estimator and the extended polynomial method. In fact with biased estimates, the extended polynomial method fails to determine the delay because none of the extended polynomial parameters will converge to zero. E x a m p l e 4.3 Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 53 Consider the system: y(t + 1) = .82y(<) + .lBu{t - 3) + e{t + 1) - 2e(t) + e(t - 1) where the input is the square wave shown in Figure 4.13. The noise sequence e(t) is a random signal ranging between —0.5 and 0.5 with equal probability. The delay estimation converges to the correct value. If the parameters of this system are to be estimated, the Approximate Maximum Likelihood estimation method can be used in a variable regression mode. 4.7 E x c i t a t i o n R e q u i r e m e n t for F M V R E The preceding analysis of the F M V R E was based on a white excitation. The auto- correlation function for the white excitation is an impulse function. In this section we consider an excitation signal with a more general auto-correlation function. In order to get good parameter estimates, the excitation signal must be rich enough. The number of different frequency components in the excitation signal should be at least half the number of unknown parameters. For example, for a first-order system having unknown gain and pole location, one sinusoid is a sufficient excitation provided that its frequency is close to the system corner frequency. If the system has an unknown delay, then a richer excitation is required if either the extended numerator polynomial or the rational approximation methods is used. In this section we prove that for such a system, the F M V R E requires only one sinusoid to estimate both the model parameters and the delay. Consider the input u(t) that has auto-correlation r ( r ) . The cross-correlation bett tween the system input and output is given by: oo r uy{r) = J2yi he{i) i=0 mpu r (r u - i) (4.54) Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 54 INPUT 5. •5. 0.0 20." _i_ 4-fl. 60= 80. WO. 120. 140. 160.., 180. 200 OUTPUT 1 0. /TA, V -5. 0.0 1 S\. i 20. 1 i A i 40. — 1 1 ^ \ V Xi V i 60. i i SO. IO0. 1 w i 120. 1— v jA I 1 a \ { r\ r 140. 160. ISO. 20Cv 140. 160. 180. 200 DELAY EST 0.0 20. 40. 60. 80. ICO. 120. SAMPLES Figure 4.13: F M V R E in presence of coloured noise. Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 55 The convolution expression in the last equation implies that the cross-correlation is generated from the auto correlation in the same way as the generation of the system output from the input. The function E\ is then generated from the input auto-correlation through the filter (1 - - )G(z). 1 Z In practice, a periodic signal is commonly used as the excitation signal, for instance, a P R B S , a square wave, a sinusoidal wave, etc. In this case, the function E\ is also periodic with the same frequency and its maximum appears repetitively. In order to obtain a unique value for the estimated delay, there must be only one maximum of E x in the range [fc ; , k \. This leads to the lower bound on the excitation signal period m n max T • T > T {k e s max — fc i ) m n (4.55) On the other hand, the frequency of the excitation signal should be fast enough to produce a sensible phase difference between the input and the output. A n upper bound on a sinusoidal excitation period is derived in the next section. 4.7.1 Sinusoidal Excitation Theorem 4.3 Using the FMVRE, the estimated delay for a minimum phase stable system that has a pole-excess of one is unbiased if the sinusoidal excitation period is related to the fastest time constant of the system r ; m T < iv^2T r e s min n by: (4.56) Proof: The F M V R E does not impose constraints on the sampling rate choice. Then, the sampling time T is independent of the delay estimation and is usually selected s to be faster than the fastest time constant of the system to provide good parameter Chapter 4. The Fixed Model Variable Regression Estimation estimation and closed loop control. 56 (FMVRE) This fast sampling is required when assuming sinusoidal input and output. Now, consider a stable minimum phase system of pole excess equal to one. If the excitation frequency iv is higher than the highest corner frequency (corresponding to e i~ in), m then the phase shift of the system is approximated by: - (TT/2 - arg(G(z)) ~ -{» T ) e d The first term in the last equation represents the phase shift due to the system time delay . The second term approximates the phase shift due to the other system dynamics. The excitation signal is sinusoidal and so its auto-correlation function is: T"u(T) = r (0) cos(u> T) u E This function has an amplitude of r (0) and is taken as our reference sinusoid, so its u phase is zero. Due to the existence of the sample and hold in the system, the sampled auto-correlation function will have a phase of one half of a sample, i.e.: -u; T,/2 arg(r ) = e u It is then direct to show that: arg((l - Tr/2-u> T = e s The phase of the function E\ is then given by: arg(£; ) 1 = arg((?(l - = -v T e d z-> ) u - u> T e s H — For correct delay estimation, the maximum of E\ should occur at the (fc + 1) th sample, which corresponds to a phase shift of arg^x) = — u) (k + 1)T, = —us (Td e e + T„). It is Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 57 then necessary that the last term in the last equation contributes a phase shift of less than one half of a sample so that the sample at which E\ is maximum does not change. The condition is then: 1 and the excitation frequency is given by: / 2 T r 5 ' mm ana since 2TT then T e < T \J2 T T t m i n and this proves the theorem. • Computer simulations show that when the upper bound of the excitation signal given here is exceeded, a correct delay estimation can still be obtained. We believe that the reason is that the sampling process of the sinusoid does not only result in a phase shift but also in a distortion of the wave. The sampled sinusoid actually has more than one frequency component which means a richer excitation. Consequently, the condition based on one sinusoid can be relaxed. E x a m p l e 4.4 This simulation example demonstrates the results of Theorem 4.3. Consider the first order system: G{s) = - —^— TiS e The system delay and time constant change with time as follows: (4.57) Chapter 4. The Fixed Model Variable Regression Estimation 58 r Time T 0 < t < 200 2 5 200 < t < 400 4 3 600 < t < 600 0 7 d (FMVRE) The system is stable and minimum phase. For this system: T„ = lsec, k max = 4, k min = 0 and T i = dsec is the fastest expected time constant. The conditions for the excitam n tion frequency are then: T > (4 - 0) = 4sec e T < 7T (2)(3)(1) = l.Wsec / e V The value of T = 6s ec is selected which satisfies the excitation requirement. e The parameters and the delay of the corresponding discrete model are shown in Figure 4.14. The results show that one sinusoid is a sufficient excitation to identify the system. 4.8 Robustness of the F M V R E The approximation of a high-order system by a low-order model usually suffers from lack of robustness. If a stable system is excited by a high frequency and the resulting system phase cannot be described by the low-order model, then the estimated model can become unstable and this can destabilize the closed-loop. The following example illustrates this fact (Astrom 1984-c). E x a m p l e 4.5 Consider a model given by: G (s) m = - ^ T s -f a (4.58) Chapter 4. The Fixed Model Variable Regression Estimation 10. ' <N/I (FMVRE) 59 DELAY EST -i r- 5. O.O 0.0 O.Q 50. 100. 150. 200. 250. 300. 350. 400. 450. 500. 550. 60< POLE EST -0.5 -1 .0 0.0 0.0 50. 100. 150. 200. 250. 3D0. 350. 400. 450. 500. 550 50. 100. 150. 200. 250. 300. 350. 400. 450. 500. 553. 60 SAMPLES Figure 4.14: F M V R E using only one sinusoid. Chapter 4. The Fixed Model Variable Regression Estimation 60 (FMVRE) used to describe the system: G ' W = (» + l)(/ 5 + 30, 229) ( 4 + - 5 9 » The system dynamics correspond to the nominal plant 2/(5 + 1) cascaded with 229/(s + 2 30s + 229). The system thus has two poles s = —15 ± 2j, which have been neglected in the model used. If the excitation signal is sinusoidal of frequency u; then in order to match the system response at this frequency, the model should satisfy: b 458/(259 - u ; ) ju + a ;7'a; + (229 - 31a; )/(259 -u> ) 2 2 2 (4.60) then the model parameters are given by: 6 = 458/(259 - a ; ) 2 a = (229 - 31a> )/(259 - a; ) 2 2 The model describes the system accurately only at very small u>. As u> increases, we get the following estimates: 0 < IO < 2.71 a > 0 2.71 < <JO < 16.1 u; = 16.1 a < 0 stable model unstable model a = ± o o unstable model W > 16.1 a > 0 stable model It is clear that the presence of unmodelled dynamics can destabilize the controller if the excitation frequency is high enough. Introducing a variable delay in the low-order model rectifies this problem. The phase of the delay part in the model is the product of the exciting frequency and the modelled time delay, and this phase can be as large as required to match the phase of the system. The resulting model is always stable and exactly matches the system response at the exciting frequency as proven in the following theorem. Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 61 T h e o r e m 4.4 A model of the form: G (s) = m G {z) e~ < — s+ d f z m s -k bo z+a x where T is the time delay estimated by the FMVRE, is stable and exactly matches the d response of any stable system having a negative phase at a given frequency. Proof: Let the magnitude and phase of the system at the excitation frequency uj e be given by \G,{JOJ)\ and <J), respectively. From the sinusoidal excitation analysis, the phase of the function E is given by: x arg(^) = arg(G(l-*->„) = n/2-co T, + <l>, e (4.61) Since the maximum of E\ occurs at the (k + 1) sample corresponding to the shift arg( E ) = -cv (k + 1)T, = -co {f + T ), then f J 1 e e d t is given by: d -u> T - 7r/2 + <f) e d t but since T is an integer multiple of T , the exact expression of T is: d t d T,integerC J ) f = /2+ d (4.62) )t then, TT/2 + <f, - T co < -Lc T s B e e d < TT/2 + <f>, (4.63) Now let the phase resulting from the first-order part of the model be <p , then the m estimation of the model parameters should satisfy: <f>m-"ef = d (4.64) Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 62 Substituting the last equation in the one before we get: -TT/2 < <£M < -TT/2 + 2 > (4.65) e which means that: 0 < a/w < tan(r„u; ) e (4.66) e So d is always positive which means a stable model that matches the phase of the system. To match the magnitude of the system and the model, 6 must satisfy: b=\G.(ju> )\y/u,* e + a* (4.67) which is always positive and finite and this completes the proof. • It is to be noted that the estimated delay variation required for exact frequency response matching is not large. At low frequency the phase of the system is | < (7r/2 — u> T„) and this phase is totally represented by the first-order model phase, then e the estimated delay is zero (Tj = 0,k = 0). At high frequency the phase of the system reaches its final value equal to the system pole excess times 7r/2 and the expression in Equation 4.62 will reach the value of zero. Between these two frequency extremes, k will reach a maximum given by the ratio in Equation 4.62. E x a m p l e 4.6 Now let us apply Theorem 4.4 to the system in Example 4.5. The Bode plot of the system is shown in Figure 4.15. The operating frequency is selected to be u) = 16.1rad/sec e which corresponds to a phase shift of 180°(<j) = — 7r). Example 4.5 shows that the pas rameters of the first order model are both infinitely large at this frequency. We start by selecting a sampling frequency adequate for the operating frequency, and this is taken as 10 times the operating frequency. This means that the phase shift Chapter 4. The Fixed Model Variable Regression Estimation 63 (FMVRE) MAGNITUDE PHASE too. 0.00 Lu LU or 100. CD LU Q 10 10 10 10 RAD C A N S / S E C O N D Figure 4.15: Bode plot of the third-order system 10' Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 64 due to one sample is u> T — ir/5 and T = .039sec. According to Equation 4.63, the e s s Estimated delay should be given by: (TT/2) - (TT) - (TT/5) < -(u> T,)(f /T.) e < TT/2 - (TT) d -7/IOTT < -(ir/5){f /T.) d < -TT/2 (4.68) (4.69) then: -3.5 < -(f /T.) < -2.5 d and since k = T /T d 6 = integer, then k = 3 is the only value possible. Figure 4.16 shows the excitation signal and the estimated delay that converges to the expected value of 3. The phase shift due to the model pole is given by Equation 4.65 as: (f>m = <t>, + u T e d = </>,+ = - T T + 3TT/5 = - 2 T T / 5 = — tan(w /d) u T,(T lT ) e d s e and hence the discrete model pole is given by: di = -exp(-r„d) = - exp((u; r ,)/(-a; /d)) e i e = -exp(Kr )/(tan- (</) ))) = exp((7r/5)/(tan- (-27r/5))) = -.815 1 a 1 m which is exactly the value estimated, as shown in Figure 4.17. This example showed the improvement in the model robustness when a delay is included in the model and the operating frequency is a sufficient excitation to estimate the delay. Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 65 INPUT 0.0 10. 20. 30. 40. 50. 60. 70. SO. 90. 100. UO. 120. 130. HO. 150 SAMPLES Figure 4.16: Delay estimation for the case of model order mismatch Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) POLE EST D -Q.SO -0.85 _I •1.05 fl.O 0. 06 10. -i 20. 1 30 - i 40. ' L. 50. ~~"7" 60. -i 70. 1 80. 1 90. 1 100. 110. I2fl r -i 0.02 0.0 0 0 10 20 30. 40. 50. 60. 70. 80. 90. 100. 110. 120 SAMPLES Figure 4.17: Parameter estimation for the case of model order mismatch 1 r- Chapter 4. The Fixed Model Variable Regression Estimation 4.9 67 (FMVRE) Systems with fractional delay By using the standard Z transform in all of the preceding analysis, we assume that the system delay (T ) is an integer multiple of the sampling time (T„). If the system d has a fractional delay, the modified Z transform (Z ) is used (Kuo, 1980). The zeros m of the sampled system change significantly if fractional delay exists, while the poles of the sampled system do not change. For example, consider the first-order system: Y(s) = - < -±—U(s) T (4.70) s e +1 TS where: T = kT. + (1 - m)T„ 0 < m < 1 d the system delay has k integer number of samples plus a fraction of a sample determined by the value of m. The sampled system is given by: »W = ^YI^'-M 1 + aq <- > 4 71 1 x where: oi = -e~ ' b = l-e- b l 0 T /T m T 'l T -mr./r _ -T./r = e e If m = 1, then b\ = 0 and b — 1 + a which corresponds to the sampled system with k 0 x samples of delay. In this case, there is no zero in the sampled model. For other values of m there will be a zero in the sampled model given by the ratio bi/b - This zero can 0 easily be outside the unit circle, which means a non-minimum phase model (Astrom et al. 1984-b). In general, the presence of fractional delay can be modelled by increasing Chapter 4. The Fixed Model Variable Regression Estimation 68 (FMVRE) the number of zeros in the model by one. In other words, the presence of fractional delay means that degB = n + 1. In the F M V R E we maximize the function Ei(k) = r & (l u + k). If k = k \ y m= is the value that maximizes E\ for integer delay, then k = fc i + (1 — m) is the value that m= maximizes E\ for fractional delay. But because the values of E\ are considered only at integer values offc,then E\ is maximum at either k = k -\ or at k — fc m m=1 -f 1 depending on the system under consideration and the value of m. This means that the estimated delay can be one step more than the value estimated for integer delay. This difficulty can be overcome by extending the numerator polynomial by one extra parameter. Let the model for integer delay be: G i{z) = z' m= -fc ' ftp*" + - - - + S - i z ~ 1 m=1 n 1 + ax*" H + a z~ 1 n n then, the model for fractional delay should be: _- b_ +b z- + .-- + b ^ z- 1 k ^m^x{z) = z 1 r — + n 0 n ; — - 1 +az x j — 1 1 — ; — + •••+ a z b z- n 1 n — n n Two parameters are added to the numerator polynomial of G =i] the last parameter m in the polynomial (b ) represents the increase in the number of zeros and the the first n parameter (fc-i) corrects the bias that may result in the delay estimation. If a bias results in the delay estimation (k — fc =i + 1) then b = 0 and 6_ ^ 0 give the m 5 n a required model. On the other hand, if k — fc i, then b ^ 0 and 6_i = 0 give the m= n required model. Not every system will require this extension of the numerator polynomial because of the delay estimation bias. A first-order systems with rich excitation will have the function r u y similar to the system impulse response. This means that E\{k) is zero for k < k and E\(k) is maximum for k = k and 0 < m < 1. The presence of fractional delay does not affect the delay estimation for first order system with rich excitation. Chapter 4. The Fixed Model Variable Regression Estimation (FMVRE) 69 The following example demonstrates the F M V R E + R L S estimation of the model in presence of a fractional delay. E x a m p l e 4.7 Consider the first-order system with fractional delay: 1 G(s) + 1' TS G(z) = -^ ^ + ^ l + a-izb 1 b z 1 The sampling time is (Td) T„ = lsec. and system time constant r = 4sec. The time delay variation and the resulting sampled system parameter are as follows: m k 1.5 .5 1 [150, 300] 2.75 .25 [300, 450] 3.5 [450, 600] [600, 750] t T [0, 150] bo h .7788 .1175 .1036 2 .7788 .0605 .1606 .5 3 .7788 .1175 .1036 4.25 .75 4 .7788 .1709 .0502 5 1 5 .7788 .2212 0 d The input used is a square wave of 40 samples period. The delay estimation is shown in Figure 4.18. The delay estimation converges to its correct value for the different value of m, which means that the square wave input is a rich excitation for the system and no need to extend the numerator polynomial. Figure 4.19 shows the estimates of the model parameters a ,b x 0 and b\. The model parameters converge to their correct values for the different values of m. 4. The Fixed Model Variable Regression Estimation (FMVRE) 70 INPUT/OUTPUT (1.0 50. 100. 150. 200. 250. 300= 350. 400. 450. 500. 550. QOd. 650. 700. 75 DELAY EST. -i 1 1 r 0. 2. 0.0 50. 100. 150. 200. 250. 300. 350. 400. 450. 500. 550. 600. 650. 700. 75 SAMPLES Figure 4.18: Delay estimation for system with fractional delay er 4. The Fixed Model Variable Regression Estimation (FMVRE) T 1 ( .0 \ ^ • 1 1 71 PARAMETER EST „ 1 1 1 1 1 » 1 I I 1 | 1 1 1 1 1 f 1 1 1 1 1 5 0 . 100.. 150. 2 0 0 . 2 5 0 . 3 0 0 . 3 5 0 . 4 0 0 . 450 . 5 0 0 . 5 5 0 . 6 0 0 . 650. 7 0 0 . 75 1 ( 1 1 i i i i i ( r i T " — r ^ l ^ s ^ 1 1 1 1 1 1 i i t ( : • r i — L 1 1 1 i i i i i 0.0 50. 100. 150. 200. 250. 300. 350. 400. 450. 500. 550. 600. 650. 700. 75 0.2 -i r -i r- 0. t 0. I 0.0 50. 100. 150. 200. 250. 300. 350. 400. 450. 500. 550. 600. 650. 700. 75 SAMPLES Figure 4.19: Parameter estimation for system with fractional delay Chapter 4. 4.10 The Fixed Model Variable Regression Estimation (FMVRE) 72 Conclusions In this chapter, a fixed model is used with the V R E , the result is the F M V R E . The F M V R E has the same advantages over the extended polynomial method and the rational approximation method as the E M V R E since no additional parameters are required to estimate the delay. However, the F M V R E does have unique properties that make it superior to the E M V R E . These advantages are: 1. The delay estimation is decoupled from the parameter estimation and consequently it always converges for first and second-order systems. 2. The delay estimation is bounded for any sampling rate. 3. The delay estimation for a high-order over-damped system is equivalent to the delay obtained by the Ziegler-Nichols method. The method can therefore be used for the auto-tuning of controllers based on the Ziegler-Nichols approximation. 4. The delay estimation does not increase the excitation richness requirements over those for good estimation of the system parameters. 5. The estimated model is stable in the presence of unmodelled dynamics and exactly matches the frequency response of the system at the operating frequency. Chapter 5 V R E - B a s e d Adaptive Control In this chapter the model parameters estimated by the RLS and the model delay estimated using the V R E are used in the controller design and the response is compared with that obtained if the design is based on the extended numerator polynomial method. Three controllers are covered in this chapter and other controllers can also be used. The controllers selected here are the Pole-Placement, the Dahlin and the Generalized Predictive controllers. They are chosen because of their practical importance and because of their capability of regulating systems with dominant time delay (Astrom 1989, Dumont 1982-b and Clarke 1988). The design is based on the certainty equivalence principle which assumes exact knowledge of the system model. Then the estimated model is used in the design instead of the actual one. In all cases, only deterministic systems are considered. Stochastic systems can be dealt with systematically as extensions to the designs given here. The controller design is indirect which means that the estimated parameters are not those of the controller. Direct controller design is also possible. The structure of the estimators and the controller is shown in Figure 5.20. The three adaptive controllers will be applied to the following system: s jw l z 0 + 2£s/w 0 + 1 The sampling rate is T, = lsec and the system parameters and delay change with time as follows: 73 Chapter 5. VRE-Based Adaptive Control Figure 5.20: Structure of the Estimators and the adaptive controller 74 Chapter 5. VRE-Based Adaptive Control 75 Time T C b w 0 < t < 100 7 1.0 0 1 100 < t < 200 4 1.5 1 1 • 200 < t < 300 1 2.0 1 1 300 < t < 400 1 2.0 -1 1 d 0 The input to the system is a square wave of a 20 samples period. The values of kmin — 0 and k max — 8 are used in the delay estimation. The system is non-minimum phase for t £ [300,400] because b is negative which means that special precautions should be taken in the control design (Astrom 1980-b and Clarke 1984). The nonminimum phase system can be written as: G(s) — GNMP GUP (5.73) wnere: -\b\a + l GNMP GMP (5.74) \b\a + l _ \\ + .-T*S b s /w 2 2 0 s 1 + 2(s/w + 1 (5.75) 0 The second-order system analysis of the last chapter showed that the pure delay term in GMP can be accurately estimated using the F M V R E . The rational function GNMP has a unity gain at all frequencies, the same as a pure delay function. The phase of GNMP 1S the same as that of the pure delay function e^ ' '*^ at low frequencies. In fact, GNMP 1S _2 b found as a fundamental building block in many delay rational function approximations, e.g. Pade, Laguerre and all-pass filters. This means that at low frequencies (u>|&| < 1) the non-minimum phase system can be approximated by a minimum phase model having delay. At higher frequencies the approximation can still be used but with a lesser delay to approximate GNMP- The F M V R E can automatically approximate 76 Chapter 5. VRE-Based Adaptive Control GNMP by a delay function provided that the phase of the second-order part of the model has a phase greater than TT/2 + wT as was shown in the last chapter. If the e phase of the second-order part is less than that, the approximating delay becomes less and the model gets closer to the actual system. For our example, the above discussion means that the F M V R E delay estimate satisfies: T < f . <T d d d + 2\b\ i.e., 1 < f d < 3, t> 300 The open loop F M V R E is shown in Figure 5.21 where the estimated delay converges to the value of 2 which agrees with the above analysis. 5.1 Adaptive Pole-Placement In a Pole-Placement controller (PP) (Astrom and Wittenmark 1980-a and Wellstead et al. 1979), the input-output A R M A model is used: Aiq-'Mt) = q-'Biq-'Mt) (5.76) Let the desired response from the reference signal u to the output be described by c the dynamics: Arniq-'Mt) = q-'B^q-^t) (5.77) A general linear controller can be described by: Riq-'Ht) = T{q- )u {t) - S{q- )y{t) l l c (5.78) Chapter 5. VRE-Based Adaptive Control 77 OPEN LOOP FMVRE O / W V W V A A j V W V W W W V >- 50. 8. 4-. 0. •4-. 100. 1 150. 1 200. I 250. 300. 350. i i i 400, n iu nu nu nu n unl n n n n n n n n I U U u U U U U u L J -s. 0.0 n n n n U U U LI L n i • i i 50. 100. 150. 2O0. 250. 300. 350. 50. 100. 150. 200. 250. 300. 350. 1 400, 8. t 6. 4. 2. 0 0.0 SAMPLE Figure 5.21: Open-loop response and F M V R E delay estimation 400 Chapter 5. VRE-Based Adaptive Control 78 Elimination of u between Equation 5.76 and 5.78 gives: BT y = ^AR + r-Bs"- ( 6 ' 7 9 ) To achieve the desired input-output relation, the following condition must hold: BT B AR + q~ BS A k m (5.80) m The denominator AR + q~ BS is the closed-loop characteristic polynomial. To carry k out the design, the polynomial B is factored as: B = B B~ (5.81) + where B + is a monic polynomial whose zeros are stable and so well damped that they can be cancelled by the regulator. zero. Since B + 0 = 1, there is no cancellation of any + is cancelled, it also factors the closed-loop characteristic polynomial. The other factors of this are A A When B and A , where A is called the observer polynomial. m 0 0 must be stable and its dynamics does not appear in the output. This gives the Diophantine equation: AR + q~ BS = A A B k 0 It follows from this equation that B + (5.82) + m divides R. Hence: R = RyB AR + q- B~S = AA k 1 0 (5.83) + (5.84) m The solution of the Diophantine equation is essentially the same as solving a set of linear equations. The solution is unique if A and B~ are relatively prime. It follows from Equation 5.80 that B~ must divide B m T = A B /B~ 0 m and that: • (5.85) Chapter 5. 79 VRE-Based Adaptive Control The pole placement design procedure can be summarized as follow: given A, B, k, A , B , A , m m 0 Step 1: Factor B as B = B B~ + Step 2: Solve R and S from the equation AR\ + q~ B~S = AA . k x Step 3: Form R = R B+ and T = 0 m A B /B~. X 0 m The control law is given by: Ru = Tu — Sy c To combine the pole placement design with the parameter and delay estimation, we simply replace A,B,k in the above design by their estimates A,B,k respectively. Because the polynomial i? is not known a-priori, it is safe not to cancel any of estimated model zeros, which means that B = 1,B~ = B. If there is no delay estimation and + the extended polynomial method is used instead of the F M V R E , the above design is still valid but k is replaced by fc ; and degB is n + k m degRi = n + k max n instead of degR max x — k i. m n This means that = n + k if the F M V R E is used. The reduction of the number of controller parameters directly reduces the number of equations to be solved and this means much less computations performed on-line. The adaptive pole placement is now applied to the system in Equation 5.72. The model used is second-order and the controller has no zero cancellation. The desired closed-loop characteristic equation has one pole corresponding to a time-constant of one sample and all other poles are at the origin of the unit circle. The adaptive P P / F M V R E response is shown in Figure 5.22 where the model delay converges to the actual system delay for the minimum-phase cases and to the value of 2 for the non-minimum phase case. The closed-loop response compares closely with the reference signal for both the minimum phase and the non-minimum phase cases. The model parameter estimates are shown in Figure 5.23. The delay estimation converges Chapter 5. VRE-Based Adaptive Control 80 in less than 20 samples while the parameter estimation converges much slower. It is expected then that when the extended numerator polynomial method is used instead of the V R E , the convergence is slow and thus the closed-loop performance is poor. This is shown in Figure 5.24 where the adaptive P P is designed based on the extended numerator polynomial method. Giving enough time for the model parameters to converge to their final values, the response using the extended numerator polynomial will be the same as that using the F M V R E except for the non-minimum phase system part where the extended numerator polynomial will eventually result in a better closed-loop response because it represents the system exactly. The main advantage of using the F M V R E in the design is the speed of getting a good closed-loop response. The E M V R E based adaptive P P also has the advantage of obtaining a good closed-loop response quickly. This is shown in Figure 5.25. The response of the adaptive P P / E M V R E may be different from that of the adaptive P P / F M V R E in the delay learnings period which is usually the first 20 samples after the delay changes. After the delay learning period, the adaptive P P / E M V R E gives the same good closed-loop response as the adaptive PP/FMVRE. 5.2 Adaptive D a h l i n Controller The Dahlin controller was originally proposed by Dahlin (1968). It can be thought of as a special case of the general pole placement design of the previous section. The closed-loop response required is that of a first-order with a gain of one and with a time delay equal to the open-loop delay. Also the controller input is the error between the actual output (y) and the desired one (u ), this means that T = S in the general pole c placement given above. Chapter 5. VRE-Based Adaptive Control 81 ADAPTIVE PP/FMVRE 400 3. A A-. 0. •4-. •8. 0.0 • • r h h ih 1 hhf i H r 1 .-i 50. 100. 150. 50. 100. 150. h t r i i ft IL L L If If \. Ir t 2O0. 250. 300. 350. AGO 200. 250. 300. 350. 400 8. 6. 4-. 2. 0 0.0 SAMPLE Figure 5.22: Adaptive P P / F M V R E response and delay estimation Chapter 5. —> i «3 I VRE-Based Adaptive Control o.o 1 82 ADAPTIVE 1 : 1 - -0.5 ~ • 1 PP/FMVRE ^ ^ ^ ^ ^ 1 1 0.0 tx: 1 1 0.3 50. I r1 0. 0.0 150. 200. 1 250. 300. t 350. 400. I • — ± -. , W 1 < «3 100. ' 1 1 1 -i - Pi n 1— | 1 50. 100. 150. 200. 250. 300. 50. 100. 150 . 200 . 250 . 300. 350- 0.5 0.2 JZ. . 1 0.0 0.3 Cn3 l 0. 0.0 Figure 5.23: Adaptive P P / F M V R E parameter estimates 350. 400. Chapter 5. VRE-Based Adaptive Control 83 ADAPTIVE PP/EXTEND. POL >- 400, S. -i ; 1 r — -I • 1 I rill L \\ h h h hih h h hh nh h' J |fJ 1J \JV p J _r j y JU |J jj |j 4-. 0. f •4-. S. 0.0 0.0 i 1 1 50. 100. • 150. i 200. i 250. 300. i 350. 400. 50. 100. 150 200. 250. 300. 350. 400 SAMPLE Figure 5.24: Adaptive PP/extended polynomial response i Chapter 5. VRE-Based Adaptive Control 8. 4-. >- O. ADAPTIVE 1 1 n n a n r A 1 ^ •4-. •8 i Q.Q 84 i 50. PP/EMVRE n o nf\ > | J i 100. « U u u \ J U U U V i 150. 20G. 250. 300. 4QQ, 350. 3. 4-. 0. L k •4-. •8, 0.0 r 50. 100. 150. 200. 150. 200. 250. 300. i r l r r 350. l r 400. <^ i 250. SAMPLE Figure 5.25: Adaptive P P / E M V R E response and delay estimation 4<?0. Chapter 5. VRE-Based Adaptive Control 85 Let Q be the required closed loop pole, then: A (q- ) = l m B (q- ) i m = l-Qq- 1 B-(q~ )A (l)/B-(l) 1 m The required closed loop response is given by: , = ,-'f^M = r ^ ^ ^ M W (5-86) The controller is then given by: «(') = | ( f c ( 0 -y(t)) = ^ m A 1 B~(1)B+ (A /A (l) m m ™ Mt) B A - kBm -y(0) (u {t)-y(t)) q- B-/B-(l)) k c (5-87) (5.88) The controller automatically has an integrator because R(l) = 0 and this property of the controller is required to compensate for set point changes. The adaptive Dahlin controller is applied to the system in Equation 5.72 with the same closed-loop requirements as in the adaptive pole placement design. The response of the adaptive D a h l i n / F M V R E is shown in Figure 5.26. The closed-loop response is good for both the minimum-phase and the non-minimum phase system. The delay estimate converges in less than 20 samples and the model parameters converge much slower than that as shown in Figure 5.27. The response of the adaptive Dahlin controller based on the extended numerator polynomial method is shown in Figure 5.28 which shows a slow closed-loop response. The response of the adaptive D a h l i n / E M V R E is also much better than that of the extended numerator polynomial design as shown in Figure 5.29. The use of VRE-based design accelerates the achievement of a good closed loop response. Chapter 5. VRE-Based Adaptive Control ADAPTIVE 86 DAHLIN/FMVRE >- 0.0 50. 100. i s . 4. 0. •4-. i 6. 4-. V h ih A u •8. 0.0 8. L 50. ,.. v 150. 250. i • I yi h h h i 100. _ , ,, 150. 1 4 (V I 1 i ^ f II 200. 1 ; 300. 00, 350 < L L if ( 1 t 250. 300. 350. 1 1 1 i r _. 1 400 ; i i i 2. 0.0 200. i i 50. 100. i 150. i 200. 250. 300. SAMPLE Figure 5.26: Adaptive D a h l i n / F M V R E response and delay estimation 350. _ 400 Chapter 5. VRE-Based Adaptive Control 87 ADAPTIVE O.O DAHLIN/FMVRE I I -0.5 <«3 0.0 Add 0.3 i ZS2 «d 0. - .3 0.0 0 0.5 l 1 1 1 0.2 1 1 i i i 100. 150. 200. 250. 300. 350. rr—' i — \ • i 1 1 0 0.3 0. < 43 50. : 50. 100. <SO. 100. 150. • 1 i i 200. 250. 300. 350= 2C>0. 250. 3GQ. 250. TV- • .3 150. SAMPLE Figure 5.27: Adaptive D a h l i n / F M V R E parameter estimates 400 . 400 Chapter 5. VRE-Based Adaptive Control 88 ADAPTIVE DAHLIN/EXTEND. POL >- 4QQ, 3. t 4-. 0. r ^ 1 L \ h h h h h in h j \4 |u [r w p r I i i .. i - •4-. •8. i 0.0 50. 100. 150. 200. 0.0 50. 100. 150. 200. 250. i r 300. h h h i V I * 350. 400, I 250. 300. SAMPLE Figure 5.28: Adaptive Dahlin/extended polynomial response 350. 400 Chapter 5. VRE-Based Adaptive Control 89 ADAPTIVE DAHLIN/EMVRE 150. 200. 250. 400. 300. 3. i l l 4. 0. -8. TIT r •4. 0.0 50. 0.0 50. 100. 100. 150. 200. 150. 200. 250. 250. 300. 350. 400, 300. 350. 400. SAMPLE Figure 5.29: Adaptive D a h l i n / E M V R E response and delay estimation Chapter 5. 5.3 90 VRE-Based Adaptive Control A d a p t i v e G e n e r a l i z e d P r e d i c t i v e controller In the Generalized Predictive Control (GPC), the objective of the controller is to minimize the following loss function (Clarke et al. 1987-a,b): C d=N J2(y(t + )-y™(t 2 J(N N ,N ) U 2 = E\ U d 2 [d=Ni Different choices of N , N , N x 2 u d=N ] £ pAu(t + d - l ) \ (5.89) u + d)) + 2 J d=l and p give rise to different schemes suggested in the literature. y (t + d) is the required future output d steps ahead and p is the control m weight. The future predictions in the loss function are estimated from the prediction model: y(t + d) = F BAu{t d + d - k - l ) + G y{t) d (5-90) - q- ) + q^G^ ) (5-91) where F and G are computed from the identity: d d 1 = A ^ F ^ X l 1 1 The future outputs can be written as: y(t + l) y{t + 2) = J 2 i A u ( < - f c ) = R Au{t + l-k) + yi(<) + y (t) 2 2 : y (t + N) = (5.92) R Au(t N + N - 1 - k) + y {t) N where R and y are given by: d d Biq-^F^q-^^Rdiq-^ + q^R^q- ) 1 Vd{t) = R (q- )u{t) + G {q- )y{t) X d X d (5.93) (5.94) The variable y (i) can be interpreted as the constrained prediction of y(t + d), under d the assumption that present and future control signals are zero. Chapter 5. VRE-Based Adaptive Control 91 Equation 5.92 can be written in the following vector form: y=RAu+y where: y = [y(t + l) ••• y(t + N)} . Au = [Au(t + 1 - k) • • • Au(t + N - y = Mt) T k)] T ••• VN(t)] T The coefficients of Rj are the first d — k + 1 terms of the impulse response of Q~ B/(AA), k which are the same as the first d — k + 1 terms of the step response of q~~ B j A. The matrix R is thus a lower triangular matrix: k r 0 n v - i 0 ••• 0 r ••• 0 0 f i v - 2 • • • ^ 0 If there is a time delay in the system {k > 0) then the first k rows of R will be zero. Also introduce the vector of desired future outputs: y m = [ y ( ' + l) ••• y (t + N)} m m T The expected value of the loss function can be written as: J = £ { ( y - y = (RAu+ y - y m ) T ( y - y m ) + Au Au} T P ) (RAu +y-y )+/9Au Au T m T m Minimization of this expression with respect to A u gives: Au = (R R4-^I)- R (y -y) T 1 T m (5.95) Chapter 5. 92 VRE-Based Adaptive Control The first component in A u is Au(t), which is the control signal applied to the system. The controller automatically has an integrator. To decrease the computations, it is possible to introduce constraints on the future control signal. For instance, it can be assumed that the control increments are zero after N u < N steps, where N is the prediction horizon in the loss function: Au(t + d- 1) = 0, d>N u This implies that the control signal is assumed to be constant after N u steps. The control law then changes to: A u = ( R ^ R x + piy^iym - y) (5.96) where R i is the N x N matrix: u r 0 r x 0 ••• 0 r • •• 0 0 Rl = I The matrix to be inverted is now of the order N x N . The output and control horizons u u can be chosen as follows: Ni: If the time delay is known, then Ni = k + 1; otherwise choose Ni — 1. N: 2 The maximum output horizon N 2 can be chosen such as N T 2 t is of the same magnitude as the rise time of the system. N : Should be at least equal to the number of unstable or poorly damped poles. u When the delay is estimated, only the output predictions for d > k + 1 are computed and also the first k rows in Rx are set equal to zero and this significantly reduces the computations of the control design. Chapter 5. VRE-Based Adaptive Control 93 The response of the adaptive G P C / F M V R E is shown in Figure 5.30, for the following parameters: JVj = jfe + l iV = le + 4 N 2 u == 1 p = y (t -)- d) m = 0.001 present reference output It is noted that the delay estimation converges faster than for the adaptive P P / F M V R E and the adaptive D a h l i n / F M V R E which means a shorter learning period.when the system changes. The model parameter estimates are shown in Figure 5.31. The response of the adaptive G P C based on the extended numerator polynomial is shown in Figure 5.32 for Ni = 1 and N = 12. Unlike the P P and the Dahlin controllers, the extended 2 numerator polynomial method will not give the same good response as the VRE-based G P C design after the parameters converge. The reason for this difference is that the design of the G P C depends on the choice of N and N . In the VRE-based design, N\ x and N 2 2 change as the estimated delay changes, which means that, after the delay, we get essentially the same closed-loop response for the different values of system delay. On the other hand, if the extended numerator polynomial method is used, N has to x be fixed at a value corresponding to the minimum expected system delay and N has 2 to be fixed at a value corresponding to the maximum expected system delay. This means that the closed-loop response gets slower as the system delay becomes smaller. If the system delay is much less than the maximum expected delay, the adaptive G P C based on the extended numerator method will give the system open-loop response. The adaptive G P C / V R E has the advantages of shorter learning period and consistency of the good closed-loop response for different system delays. The adaptive G P C / E M V R E Chapter 5. VRE-Based Adaptive Control 94 has as good a closed-loop response as the adaptive G P C / F M V R E as shown in Figure 5.33 5.4 D e l a y E s t i m a t i o n in C l o s e d - L o o p 5.4.1 D e t e r m i n i s t i c Systems When the parameter estimation is carried out in closed-loop, the problem of parameter identifiability arises (Ljung et al. 1974, Soderstrom et al. 1975, 76 and Gustavsson et al. 1977). Let the system be described by the regression model: (5.97) y(t) = *(*)© Since we have measurements from t = 1 to the present time, the following vector equation results: Y{t) = $(<)© (5.98) where: Y(t) = [y(0 v(*-i) - y(i)f (5.99) *(0 (5.100) and the Least-Squares parameter estimate is given 0 = (* *) T The matrix ( $ $ ) T _ 1 * Y T (5.101) is called the excitation matrix and it should be invertible which means that the columns of $>(t) should be linearly independent. In open-loop estimation this means that the system input should be persistently exciting. In closed-loop er 5. VRE-Based Adaptive Control 95 ADAPTIVE 0.0 50. 100. GPC/FMVRE 150. 200. 250. 300. 350 . 150. 200. 250. 300. 350. 150. 200. 400. 5. 4-. 0. -4-. - s . 0.0 400. 8. 6. 4. 2. 0 0.0 50. 100. 250. SAMPLE n£ 300. Figure 5.30: Adaptive G P C / F M V R E response and delay estimation 350. 400 96 Chapter 5. VRE-Based Adaptive Control ADAPTIVE <«3 100. 0.3 CN2 150. GPC/FMVRE 200. 250. 300. 1 1 1 1 t " 1 ^ ^ 1 1 1 1 1 1 350. 400, 1 ( 0. <r0 0.0 t\ 50. 100. 150. 2O0. 250. 300. 50. 100. 150. 200. 250. 300. 50. 100. 1 350. 400 0.5 -° 1 <X2 | 0.2 •. 1 0.0 0 350. 400, 0.3 • 03 <J=> 0 0.0 150. 200. 250. 200. SAMPLE Figure 5.31: Adaptive G P C / F M V R E parameter estimates 350. 400. Chapter 5. VRE-Based Adaptive Control 97 ADAPTIVE GPC/EXTEND. POL 4-. >•4-. •8. O.Q SO. 100. 150. 200. 250. 300. 350. 250. 300. 350. 300. 350. 400, 8. 4-. P 0. -4-. •8. O.Q 50. 100. 150. 200. G.O 50. 100. 150. 200. 250. SAMPLE Figure 5.32: Adaptive GPC/extended polynomial response 400, 4QQ Chapter 5. S. VRE-Based Adaptive Control I 98 ADAPTIVE 1 1 •• • 1 GPC/EMVRE 1— 1 1 1 250. 300. 1 '— T 4-. >- O. •4-. •8. 0.0 J U U u U g l j y u ^ 1 50. 1 100. 1 1 150. 2O0. 1 350. 400 3. 4-. 0. r -4-. -S. '0.0 V k 50. 100. r If 150. 200. 150. 200. 250. 300. 250. SAMPLE Figure 5.33: Adaptive G P C / E M V R E response and delay estimation 350. r 400. 400 99 Chapter 5. VRE-Based Adaptive Control estimation the persistent excitation is required, but also u(t) should not be linearly dependent only on the other parameters in $(i) through the feedback path. This imposes constraints on the controller structure and if these constraints are violated a non-unique parameter estimate results. The delay estimation using V R E does not include the inverse of the excitation matrix. The condition that the excitation matrix should be invertible can be violated and still we get a unique value of delay estimation. However, as was seen in the previous chapter, a certain excitation level is required for an unbiased delay estimation. The excitation required for an unbiased delay estimation is not the same as that required for unbiased parameter estimation. A n excitation level as low as one sinusoid can be enough for unbiased delay estimation for any system order as long as the system pole excess is one, while one sinusoid is not persistently exciting if the system has more than two parameters. The closed-loop delay estimation in the previous section shows that the delay does not lose identifiability as we close the loop and the estimated delay is exactly the same as that obtained by open-loop estimation. 5.4.2 Stochastic Systems If process noise exists, the output can be expressed as in Equation 4.49: y(t) = Giq-^uit) = y + G (q-')e(t) e + u y e and the input-output cross-correlation is giving by: r {t) = r {t) + r (t) uy uyu uyc The term r , represents the cross-correlation between the present input and the future uy< process noise. In open-loop estimation, the input is not correlated with the process Chapter 5. 100 VRE-Based Adaptive Control noise and hence we get r = 0, which means that the process noise has no effect on uyc the delay estimation. In closed-loop estimation, the input is indeed correlated with the past process noise through the feedback path. For input given by: R( - )u(t) = 1 q T(q- )u (t)-S( - )y(t) 1 1 c q the system input in closed loop is given by: GS/T U GS/R = 1 + GS/R < U + . 1 + GS/R < V ( 5 ' 1 0 2 ) u is not correlated with y and hence it does not contribute to the term r . The c e uyc present input is related to the past y by the closed-loop transfer-function. This means e that the present y is related to the future u by the closed-loop transfer-function, the e term r yell can then be written as: =rflb " r The function r VeU w (5103) is generated from the auto-correlation of the process noise r through Vc the closed-loop transfer function and then the function r uyc image of r VcU can be obtained as a mirror around the zero shift axis (r (t) — r (—t)), this is explained in Figure uye 5.34. The function r (t) uyr yeU does not, in general, have a zero value for t > 2fc ;„, which m means that we may, or may not, get a bias in the delay estimation because of the non-zero value of r uy< ,(i). If the process noise y has a non-zero auto-correlation for only t < 2fc {„, then e m r .(t) = 0,t £ [k i , k ] and no bias in the delay estimation results. A white process uy< m n max noise is one special case where y has a non-zero auto-correlation only for t = 0 and we e conclude that such a process noise does not affect the delay estimation in closed-loop. The closed-loop system does not have to be stable to apply this conclusion. If the closed-loop is unstable, r (t) will grow indefinitely, but not r (t). ycU uy<: Chapter 5. 101 VRE-Based Adaptive Control 20. NO[SE CORRELATION EUNCTIONS 1 — 1 1 I 1 IO. ^ 1 \ 1 1 1 ' 1 \ 1 1 1 6. 8. 11 •' ] v. ''O.Q -10. •12. 1 1 -10. -8. 1 •6. 1 1 -4. -2. 0.0 20. I I -i 10. ! 1 2 . r 110. S5 4- "-mtn-^ N 0.0 -10. •12. -ID. -8. -8. -4. 6. 0.0 20. 1 10. V -10. -8. 4. 10. ! 1 2 . 8. 10. 1 2 . T tn-*j O.O -10. •12. 8. -2. 0.0 2. TIME SHIFT 4. Figure 5.34: Correlation functions of the process noise 6. Chapter 5. 102 VRE-Based Adaptive Control If the process noise does not satisfy the above condition on its auto-correlation, a delay estimation bias may result and one of the following options may be taken to eliminate this problem: • If the process noise has known properties in the bandwidth of interest, the input and output signals can be filtered through the inverse of the estimated noise filter before the estimation is carried out. This approach is similar to that taken by Clarke et al. (1987-a,b) in the design of the G P C . In this case: (G^y(t)) = y {t) = s G(G; u(t)) + l Gu (t) G;*G e(t) e + f y (5.104) e yf and Uf are the signals used in delay estimation. The filtered process noise y e should now satisfy the required condition on its auto-correlation. • If the process noise has unknown properties we can replace the input signal u(t) used in the delay estimation by the command signal u (i) and an unbiased delay c estimation should result. This is based on the fact that the delay of the open-loop system equals the delay of the closed-loop system (no delay is arbitrary inserted in the controller). The control variable u is uncorrelated with y and hence the c e process noise has no effect on the delay estimation. This method is equivalent to operating the V R E in an auto-tuning mode instead of the self-tuning mode. • E M V R E can be used with the Approximate Maximum Likelihood to estimate the system parameters and the noise filter parameters. 5.5 Conclusion In this chapter we showed how to utilize the direct delay estimation obtained from the V R E to design controllers with the minimum number of parameters. The closed-loop Chapter 5. VRE-Based Adaptive Control 103 response is much better than that obtained without the V R E . The design of V R E based Pole Placement, Dahlin and Generalized Predictive Controllers is given. The direct delay estimation from the V R E allowed us tune the G P C and get the same closed-loop performance for any system delay, this was not possible otherwise. Unlike the system parameters, the delay cannot not lose identifiability in closedloop identification. For stochastic systems, a bias in the delay estimation may result. This bias can be avoided by filtering the input and output signals through the inverse of the noise filter, if known. If the noise filter is not known, the delay estimation can be carried out in the auto-tuning mode or by estimating the noise filter parameters using the Approximate Maximum Likelihood. Chapter 6 E x p e r i m e n t s on D e l a y E s t i m a t i o n In order to demonstrate the efficiency of the new delay estimation method, two experiments are carried out. In the first experiment, the F M V R E is used to estimate the model of a temperature control experiment. The estimated model is used in the adaptive Dahlin controller design. The experiment demonstrates the efficiency of the F M V R E in closed-loop adaptive control. In the second experiment, the E M V R E is used to estimate the delay of a pulp refiner industrial process. The experiment demonstrates the efficiency and speed of the E M V R E with noisy measurements. 6.1 Temperature Control Experiment The schematic diagram of the experimental setup is shown in Figure 6.35. A similar setup was built by Gendron et al. 1990-a, b. The setup consists of a tank where cold water is heated and then passed through a series of three long hoses. The water temperature at each hose end is measured and can be used as the process output. The process input is the electrical energy delivered to the 110V, 15Q0W heater through the electronic power switch (Triac). The process is controlled by an IBM — PC compatible computer and the software is written in Microsoft Quick Basic. The cold water comes from the domestic water line. As expected, the temperature and the flow rate of this water change significantly with time. The change in the cold water temperature is a disturbance to the process. Flow rate changes directly change the process dynamics and delay. The hoses used are simple 1/2' ', 50' garden hoses. 104 Chapter 6. Experiments on Delay Estimation 105 Input Flow Output Flow Temperature Measurement Figure 6.35: Temperature control experiment 106 Chapter 6. Experiments on Delay Estimation There is a significant heat loss from the hoses due to convection. This is another disturbance to the system. The sampling period used is 42.5sec. and this results in a time delay of about four samples in every hose. The maximum value for the delay search was set to 15 samples. The gating circuit turns-on the Triac for an integer number of the power supply cycles and the duty cycle of the Triac is controlled from 0% to 100%. The computer controls the duty cycle through an 8-bit output port. The temperature is measured by the LM335A solid state devices which provide linear temperature measurements to the computer through the A/D ports. The model used is a second-order model of the form: y(t + 1) = -e y{t) - e y{t - 1) + 8 u{t - k) + e u{t - 1 - k) + 0 x 2 3 4 5 The bias term 6 is used to represent the operating point changes due to the cold water Z temperature changes and the convection heat loss through the hoses. The parameters Q to 0 are estimated using RLS and k is estimated using F M V R E . x 5 Figure 6.36 shows the experimental results over a period of 1600 samples i. e. about 19 hours. The controlled output is switched to a different temperature sensor about every 400 samples. The controlled output is required to track a square wave variation in the reference signal. Since we are mainly interested in delay estimation results, the adaptive Dahlin controller is used. A more complicated controller will result in a better closed-loop response using the same estimation model. The results show excellent convergence and speed for the delay estimation. As the estimated delay gets very large, it can oscillate between two adjacent values instead of converging to one of them. The reason is that the function E becomes flatter near its x maximum if the excitation signal is not rich enough or if the forgetting factor is small. However, both delay values represent the process reasonably well, keeping in mind that 107 Chapter 6. Experiments on Delay Estimation OUTPUT 30. 25. 20. Q.QQ 200= 400. 600. 80G. 1000. 1200. 1400. 1000. 1200. 1400. 1600 INPUT i .<? 0.5 0.00 200. 400. 600. 800. EST 0 16QC DELAY 15. nn—r to. 5. 0. 0.00 200. 400. 600. 800. 1000. 1200. SAMPLES Figure 6.36: Delay estimation and temperature variation 1 4 0 0 . 1SOC Chapter 6. Experiments on Delay Estimation 108 any alternative method would require many more parameters to represent the process for such a large value of the delay. The controlled temperature shows some offset in the first 400 samples in Figure 6.36. The Dahlin controller has an integrator which should be able of eliminating this offset. The period of the reference square wave signal is not long enough to let the system reach its steady-state value. After the first 400 samples, the estimator has a better model of the system and it eliminates the offset more efficiently. The input signal shows significant changes in its average values over the experiment duration. This is required to counteract the changes in the heat loss by convection and the changes in the ambient temperature. It is important to note that the output average value does not change over the experiment duration. 6.2 Reject P u l p Refiner E x p e r i m e n t The use of chip and pulp refiners is spreading in the wood pulp industry, especially in Thermo-Mechanical Pulping ( T M P ) (Dumont et al. 1982-c and 1988). A reject pulp refiner consists of two counter-rotating grooved plates, with pressure exerted on one of them by a hydraulic cylinder as in Figure 6.37 (Dumont 1982-a). Wood pulp and dilution water are fed near the axis and forced to move outward between the plates by centrifugal and frictional forces. Steam and pulp are discharged at the periphery. The specific energy, or energy per mass unit of wood fibers, is a major factor controlling the pulp quality. This means that the feed-rate and the motor load must be controlled. To control the motor load, the plate gap is manipulated by adjusting the hydraulic pressure. The process input is the gap between the two plates and the process output is the motor load in MW. A generated P R B S is added to the gap reference setting. The sampling time is bsec. and the duration of the pulse is sent to a micro-dial which sends Chapter 6. Experiments on Delay Estimation PULP P L A T E S REFINED P U L P Figure 6.37: Schematic representation of a pulp refiner. 109 Chapter 6. Experiments on Delay Estimation 110 a signal to a servo valve directing oil to the cylinder. The input and output measurements are shown in Figure 6.38 and the model delay and gain estimation are shown in Figure 6.39. The delay estimation converges to the value of one in less than 20 samples, which is extremely fast for this noisy measurements. The process model is known a-priori to be pure gain plus delay and the delay is in the range of k 6 [0,5]. Because the exact process gain and delay are not known, the E M V R E delay estimation is compared with that obtained from off-line analysis of the same measurements. The input auto-correlation and the input-output cross-correlation functions are shown in Figure 6.40. There is a shift of two samples between the first non-zero auto-correlation value and the first non-zero cross-correlation value, assuming that the threshold of zero and non-zero is 10% of the maximum value obtained. This means a delay of one sample, i.e. y{t + l) = fc„u(<-1). and this value agrees with the E M V R E estimation. 6.3 Conclusions In this chapter we presented experimental results for closed-loop and open-loop delay estimation. Adaptive D a h l i n / F M V R E is used to control the temperature in an ex- periment having long delay. The results show fast adaptation and good closed-loop performance over the wide range of delay variation. In the second experiment E M V R E is used to model the Reject Pulp Refiner. Although the measurements are very noisy, the delay estimation converges in less than 20 samples to the same value obtained by off-line analysis. The experiments demonstrate the applicability of the V R E delay estimation methods. 111 er 6. Experiments on Delay Estimation 5. 4 PULP REFINER j . 1 1 1 1 ^ > ' 5.2 g MEAS. ' r ' 1 ft I I i Q 0 20 I 1 1 : — i 1 1 ' 1 1 ' ' ' ' 4-0. 60 . 80. 100. 120. 140. 160. 180 . 200 . 220 . 240. 260. 280. 30 SAMPLES Figure 6.38: Measurements on the pulp refiner 112 Chapter 6. Experiments on Delay Estimation PULP REFINER EST. ,0 20. + 0. 60. 80. 100. 120. 140. 160. 180. 200. 220. 240. 260. 280. 300. 02 <t3 0.2 0.0 20 40 60. 80. 1<?0. 120. 140. 160. 180 . 200 . 220 . 240. 260. 280. 300. SAMPLES Figure 6.39: Delay and gain estimation of the pulp refiner Chapter 6. Experiments on Delay Estimation 60. 113 PULP REFINER CORRELATION FUNCTIONS 40. t? 2 0 . s s 0.0 •20. •10. s ^ ixi D -6. -4. 1ST" -czr TSZT •2. 2. 0.0 4. 8. 20. N 15. 5» 10. N S N S o.o •5. -10. N \ "ET* -4. -2. 0.0 TIME 2. SHIFT Figure 6.40: Correlation functions of the pulp refiner measurements N o. Chapter 7 Conclusions 7.1 S u m m a r y of the W o r k In this thesis we have discussed new methods for direct estimation of the unknown delay of a system utilizing as much as possible the information available a-priori. We developed the Variable Regression Estimation which estimates the system delay interactively with the other process parameters (the Estimated Model Variable Regression Estimation) or which estimates the system delay independently of the process parameters (the Fixed Model Variable Regression Estimation). The theoretical study, computer simulations and experimental results show the superior properties of the proposed methods over the traditional delay estimation methods. The results of the thesis are summarized as follows: • The Variable Regression is introduced and joined with standard parameter estimators. The parameter vector has exactly the same number of unknowns as in the known delay case. No additional parameters are used to represent the delay. • The Estimated Model Variable Regression Estimation is introduced for systems of small range of parameter variations. The method has only one convergence point corresponding to the correct delay. Coloured noise does not affect the delay estimation. 114 Chapter 7. 115 Conclusions • The Fixed Model Variable Regression Estimation is introduced for systems of unknown parameters. The estimated delay is accurate for first and second-order systems. The estimated delay approximates the high order dynamics for higher-order systems and the estimated delay is close to that obtained by the Ziegler-Nichols approximation. The F M V R E can also be applied to systems with fractional delay. • Proofs of the F M V R E properties are derived. The estimated delay is consistent, bounded and unbiased even in the presence of coloured noise. The estimated model is stable and matches the system frequency response at a given frequency in the presence of unmodelled dynamics. The excitation richness required for correct delay estimation is less than the excitation richness required to estimate the other process parameters. estimation. One sinusoid can be enough for correct delay For known parameters system, any non-steady excitation signal is enough for correct delay estimation. • New adaptive versions of the pole placement, Dahlin and generalized predictive controllers are presented. The controller designs are based on the V R E which results in minimum number of the controller parameters, faster adaptation and good closed-loop response. The adaptive G P C / V R E has the same performance for different system delays. This is not the case without direct delay estimation. • Unlike the other model parameters, the delay cannot loose identifiability in closedloop. For stochastic systems, a bias may occur in closed-loop delay estimation depending on the noise and the system structure. This can be eliminated by filtering the input and output signals or estimating the delay between the output and the reference signal. Chapter 7. Conclusions 116 • Experimental results for closed-loop and open-loop delay estimation are presented. Adaptive D a h l i n / F M V R E is used to control the temperature in an experiment having long delay. The results show fast adaptation and good closed loop : performance over the wide range of delay variation. In the second experiment E M V R E is used to model the Reject Pulp Refiner. Although the measurements are very noisy, the delay estimation converges in less than 20 samples to the same value obtained by off-line analysis. The experiments demonstrate the applicability of the V R E delay estimation methods. 7.2 F u t u r e R e s e a r c h Directions Further research can be directed to investigate the properties of the E M V R E . This should study the delay estimation equations and the parameter estimation equations simultaneously. The E M V R E behavior with high-order systems is not fully understood yet. Further research can also be directed to study the delay estimation in closed-loop and derive the condition on the noise required to guarantee unbiased delay estimation. Having done this, the V R E can be used in the design of controllers for stochastic systems such as minimum-variance controllers. We believe that the VRE-based adaptive controllers can provide excellent control for linear deterministic systems. Applying the developed techniques of estimation and adaptive control to industrial processes having unknown or varying delay is our next target. Bibliography Allison B . J . , G . A . Dumont, L . Novak and B. Cheetham, Adaptive Predictive Control of Kaymer Digester Chip Level Using Strain Gauges Level Measurements, A I C h E Annual Meeting, San Francisco, 1989 Astrom, K . J . and B. Wittenmark (1980-a). Self Tuning Regulators Based on Pole-Zero Placement. I E E proc. part D 127: 120-130. Astrom, K . J . (1980-b). Direct Methods for Non -minimum Phase Systems. Proc. 19th C D C , pp. 611-615. Astrom, K . J . and B . Wittenmark (1984-a). Computer Controlled Systems. Theory and design. Prentice-Hall, Inc., Englewood Cliffs, N . J . Astrom, K . J . , P. Hagander and J . Sternby (1984-b). Zeros of Sampled Systems. Automatica, vol. 20, no. 1, pp. 31-38. Astrom, K . J . (1984-c). Interaction Between Excitation and Unmodelled Dynamics in Adaptive Control. Proc. 23rd I E E E C D C , Las Vegas, pp. 1276-1281. Astrom, K . J . and T . Hagglund (1988). Automatic Tuning of PID Controllers. Inst..Soc. of America, Research Triangle Park, N . C . Astrom, K . J . and B . Wittenmark (1989). A D A P T I V E C O N T R O L . AddisonWesley Co., N . Y . Biswas, K . K . and G . Singh (1985). Identification of Stochastic Time Delay Systems. I E E E Trans. A . C . , AC-23, 504-505. 117 Bibliography 118 Chandra P., C . V . Hahn, H . Unbehauen and U . Keuchel (1985). Adaptive Control of a Variable Dead Time Process with an Integrator. IFAC Workshop on Adaptive Control of Chemical Processes, Frankfurt, pp. 71-75. Chien, I. L . , D. E . Seborg and D. A . Mellichamp (1984). A Self- Tuning Controller for Systems with Unknown or Varying Time Delay. Proc A C C , pp. 905-912. Clarke, D. W . (1984). Self-Tuning Control of Nonminimum-Phase Systems. Automatica, vol. 20, no. 5, pp. 501-517. Clarke, D. W . , C . Mohtadi and P. S. Tuffs (1987-a). Generalized Predictive Control Part 1, The Basic Algorithm. Automatica, 23, pp. 137-148. Clarke, D. W . , C . Mohtadi and P. S. Tuffs (1987-b). Generalized Predictive Control Part 2, Extensions and Interpretations. Automatica, 23, pp. 149160. Clarke, D. W . (1988). Application of Generalized Predictive Control to Industrial Processes. I E E E Control Sys., vol 8, no 2, pp. 49-54. Clough, D. E . and S. J . Park (1985). A Novel Dead Time Adaptive Controller. IFAC Workshop on Adaptive Control of Chemical Processes, Frankfurt, pp. 21-26. Dahlin, E . B. (1968). Designing and Tuning Digital Controller. Instruments and Control Systems, vol. 41, pp 77-83. De Souza, C . E . , G . C . Goodwin, D. Q. Mayne and M . Palaniswami (1988). A n adaptive control algorithm for linear systems having unknown time delay. Automatica, 24, 327-341. Dumont, G . A . (1982-a). Self Tuning Control of a Chip Refiner Motor Load. Bibliography 119 Automatica, vol. 18, no. 3, pp 307-314. Dumont, G . A . (1982-b). Analysis of the Design and Sensitivity of the Dahlin Regulator. Canadian Chem. Eng. Conf., Vancouver, Canada. Dumont, G . A . , N . Legault and J . Rogers (1982-c). Computer Control of a T M P Plant. Pulp and Paper Canada, v 83, no 8, pp 224-229 Dumont, G . A . and K . J . Astrom (1988). On Wood Chip Refiner Control. Control Sys. mag. v 8, no 2, pp 38-43 Gabay, E . and S. J . Merhav (1976). Identification of Linear Systems with Time Delay Operating in a Closed Loop in the Presence of Noise. I E E E Trans. A . C . , AC-21, 711-716. Gauss, K . F . (1809). Theoria motus corporum coelestium. English translation: Theory of the Motion of the Heavenly Bodies. Dover, New York (1963). Gawthrop, P. J . and K . W . L i m (1985). Identification of Time Delay Using a Polynomial Identification Method. Syst. Control Lett., 5, 267-271. Gendron, S. (1990-a). Improving the Robustness of Dead-Time Compensation for Plants with Unknown or Varying Delay. 40th Canadian Chemical Engineering conference, Halifax, NS, pp. 138-156. Gendron, S. and A . P. Holko (1990-b). Simple Adaptive Digital Dead-Time Compensators for Low-Order SISO processes. 9Th IFAC symposium on Identification and System parameter Estimation. Gerry, J . P., E . F . Vogel and T . F . Edgar (1983). Adaptive Control of a Pilot Scale Distillation Column. Proc. A C C , pp. 819-824. Goodwin, G . C. and K . S. Sin (1984). Adaptive Filtering, Prediction and Control. Prentice-Hall, Englewood Cliffs, New Jersey. Bibliography 120 Gustavsson, I., L . Ljung and T . Soderstrom (1977). Identification of Processes in Closed Loop - Identifiability and Accuracy Aspects. Automatica, v 13, pp 59-75 Knowles, G . and E . Emre (1985). A TchebychefT Approximation Approach to the Computation of Stabilizing Compensation for Systems with Delay. Proc. 24th I E E E C D C , Ft. Lauderdale, F L , pp. 1453-1454. Kuo, Benjamin (1980). Digital Control Systems. Holt, Rinehart and Winston, Inc. Kurz, H . and W . Geodeck (1981). Digital Parameter Adaptive Control of Process with Unknown Dead-Time. Automatica, 17, 245-252. Lammers, H . C . and H . B. Verbruggen (1985). Simple Self Tuning Control of Process with a Slowly Varying Time Delay. I E E Control Conf., Cambridge. Ljung, L . , I. Gustavsson and T . Soderstrom (1974). Identification for Linear Multivariable Systems Operating Under Linear Feedback Control. I E E E Trans. AC-19, pp 836-840 Ljung, L . and T . Soderstrom (1987). Theory and Practice of Recursive Identification. M I T press, 1987. Plackett, R. L.(1950). Some theorems in least squares. Biometrika, vol. 37, pp. 149-157. Rao, G . P. and L . Siuakumar (1979). Identification of Time Lag System Via. Walsh Functions. I E E E Trans. AC-24, pp 806-808. Rao, G . P (1983). Piecewise Constant Orthogonal Functions and their Application to Systems and Control. Springer, New York. Bi bliography 121 Robinson, W . R. and A . C . Soudack (1970). A Method for the Identification of Time Delay in Linear Systems. I E E E Trans. A . C . , AC-15, 97-101. Soderstrom, T . , I. Gustavsson and L . Ljung (1975). Identifiability Conditions for Linear Systems Operating in Closed Loop. Int. J . Control 21, pp 243-255. Soderstrom, T . , L . Ljung and I. Gustavsson (1976). Identifiability Conditions for Linear Multivariable Systems Operating Under Feedback. I E E E Trans. AC-21, pp. 837-840. Wellstead, P . E . , J . M . Edmunds, D . Prager and P. Ianker (1979). Self Tuning Pole/Zero Assignment Regulators. Int. J . Control, 30:1-26. Vogel, E . F . and T . F . Edgar (1980). A n adaptive Dead-Time Compensator for Process Control. Proc. A C C . Zervos, C . (1988-a). Adaptive Control Based on Orthonormal Series Representation. Ph.D. Thesis, Dept of Elect. Eng., Univ. of British Columbia. Zervos, C . and G . Dumont (1988-b). Deterministic Adaptive Control Based on Laguerre Series Representation. Int. J . Control, v 48, no 6, pp 2333-2359. Zervos, C , G . Dumont and G . Pageau (1990). Laguerre-Based Adaptive Control of pH in an industrial Bleach Plant Extraction Stage. Automatica, v 26, no 4, pp. 781-787.
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Variable regression estimation of unknown system delay
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Variable regression estimation of unknown system delay Elnaggar, Ashraf 1990
pdf
Page Metadata
Item Metadata
Title | Variable regression estimation of unknown system delay |
Creator |
Elnaggar, Ashraf |
Publisher | University of British Columbia |
Date Issued | 1990 |
Description | This thesis describes a novel approach to model and estimate systems of unknown delay. The a-priori knowledge available about the systems is fully utilized so that the number of parameters to be estimated equals the number of unknowns in the systems. Existing methods represent the single unknown system delay by a large number of unknown parameters in the system model. The purpose of this thesis is to develop new methods of modelling the systems so that the unknowns are directly estimated. The Variable Regression Estimation technique is developed to provide direct delay estimation. The delay estimation requires minimum excitation and is robust, bounded, and it converges to the true value for first-order and second-order systems. The delay estimation provides a good model approximation for high-order systems and the model is always stable and matches the frequency response of the system at any given frequency. The new delay estimation method is coupled with the Pole Placement, Dahlin and the Generalized Predictive Controller (GPC) design and adaptive versions of these controllers result. The new adaptive GPC has the same closed-loop performance for different values of system delay. This was not achievable in the original adaptive GPC. The adaptive controllers with direct delay estimation can regulate systems with dominant time delay with minimum parameters in the controller and the system model. The delay does not lose identifiability in closed-loop estimation. Experiments on the delay estimation show excellent agreement with the theoretical analysis of the proposed methods. |
Subject |
Estimation theory Adaptive control systems Regression analysis |
Genre |
Thesis/Dissertation |
Type |
Text |
Language | eng |
Date Available | 2011-01-24 |
Provider | Vancouver : University of British Columbia Library |
Rights | For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use. |
DOI | 10.14288/1.0065633 |
URI | http://hdl.handle.net/2429/30804 |
Degree |
Doctor of Philosophy - PhD |
Program |
Electrical and Computer Engineering |
Affiliation |
Applied Science, Faculty of Electrical and Computer Engineering, Department of |
Degree Grantor | University of British Columbia |
Campus |
UBCV |
Scholarly Level | Graduate |
Aggregated Source Repository | DSpace |
Download
- Media
- 831-UBC_1991_A1 E46.pdf [ 5.36MB ]
- Metadata
- JSON: 831-1.0065633.json
- JSON-LD: 831-1.0065633-ld.json
- RDF/XML (Pretty): 831-1.0065633-rdf.xml
- RDF/JSON: 831-1.0065633-rdf.json
- Turtle: 831-1.0065633-turtle.txt
- N-Triples: 831-1.0065633-rdf-ntriples.txt
- Original Record: 831-1.0065633-source.json
- Full Text
- 831-1.0065633-fulltext.txt
- Citation
- 831-1.0065633.ris
Full Text
Cite
Citation Scheme:
Usage Statistics
Share
Embed
Customize your widget with the following options, then copy and paste the code below into the HTML
of your page to embed this item in your website.
<div id="ubcOpenCollectionsWidgetDisplay">
<script id="ubcOpenCollectionsWidget"
src="{[{embed.src}]}"
data-item="{[{embed.item}]}"
data-collection="{[{embed.collection}]}"
data-metadata="{[{embed.showMetadata}]}"
data-width="{[{embed.width}]}"
async >
</script>
</div>
Our image viewer uses the IIIF 2.0 standard.
To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0065633/manifest