Adaptive Control Of A General Class Of Finite Dimensional Stable LTI Systems
We consider the problem of Adaptive Control of finite-dimensional, stable, Linear Time Invariant (LTI) plants. Amongst such plants, the subclass regarding which an upper bound on the order is not known or which are known to be nonminimum phase (zeros in the unstable region) pose formidable problems in their own right. On one hand, if an upper bound on the order of the plant is not known, adaptive control usually involves some form of order estimation. On the other hand, when the plant is allowed to be either minimum phase or nonminimum phase, the adaptive control problem, as is well-known, becomes considerably-less tractable. In this study, the class of unknown plants considered is such that no information is available on the upper bound of the plant order and, further, the plant may be either minimum phase or nonminimum phase. Albeit known to be stable, such plants throw myriads of challenges in the context of adaptive control. Adaptive control involving such plants has been addressed  in a Model Reference Adaptive Control (MRAC) framework. There, the inputs and outputs of the unknown plant are the only quantities available by measurement in terms of which any form of modeling of the unknown plant may be made. Inputs to the reference model have been taken from certain restricted classes of bounded signals. In particular, the three classes of inputs considered are piecewise continuous bounded functions which asymptotically approach • a nonzero constant, • a sinusoid, and • a sinusoid with a nonzero shift. Moreover, the control law is such that adaptation is carried out at specific instants separated by progressively larger intervals of time. The schemes there have been proved to be e-optimal in the sense of a suitably formulated optimality criterion. If, however, the reference model inputs be extended to the class of piecewise continuous bounded functions, that would compound the complexity of the adaptive control problem. Only one attempt  in adaptive control in such a setting has come to our notice. The problem there has been tackled by an application of the theory of Pade Approximations to time moments of an LTI system. Based on a time moments estimation procedure, a simple adaptive scheme for Single-Input Single-Output (SISO) systems with only a cascade compensator has been reported. The first chapter is essentially meant to ensure that the problem we seek to address in the field of adaptive control indeed has scope for research. Having defined Adaptive Control, we selectively scan through the literature on LTI systems, with focus on MRAC. We look out in particular for studies involving plants of which not much is known regarding their order and systems which are possibly nonminimum phase. We found no evidence to assert that the problem of adaptive control of stable LTI systems, not necessarily minimum phase and of unknown upper bound on the order, was explored enough, save two attempts involving SISO systems. Taking absence of evidence (of in-depth study) for evidence of absence, we make a case for the problem and formally state it. We preview the thesis. We set two targets before us in Chapter 2. The first is to review one of the existing procedures attacking the problem we intend to address. Since the approach is based on the notion of time moments of an LTI system, and as we are to employ Pade Approximations as a tool, we uncover these concepts to the limited extent of our requirement. The adaptive procedure, Plant Command Modifier Scheme (PCMS) , for SISO plants is reported in some detail. It stands supported on an algorithm specially designed to estimate the time moments of an LTI system given no more than its input and output. Model following there has been sought to be achieved by matching the first few time moments of the reference model by the corresponding ones of the overall compensated plant. The plant time moment estimates have been taken to represent the unknown plant. The second of the goals is to analyze PCMS critically so that it may serve as a forerunner to our work. We conclude the chapter after accomplishing these goals. In Chapter 3, we devise a time moment estimator for SISO systems from a perspective which is conceptually equivalent to, yet functionally different from, that appropriated in . It is a recipe to obtain estimates of time moments of a system by computing time moment estimates of system input and output signals measured up to current time. Pade approximations come by handy for this purpose. The lacunae exposed by a critical examination of PCMS in Chapter 2 guide us to progressively refine the estimator. Infirmities in the control part of PCMS too have come to light on our probing into it. A few of these will be fixed by way of fabricating two exclusively cascade compensators. We encounter some more issues, traceable to the estimator, which need redressal. Instead of directly fine-tuning the estimator itself, as is the norm, we propose the idea of 'estimating' the lopsidedness of the estimator by using it on the fully known reference model. This will enable us to effect corrections and obtain admissible estimates. Next, we explore the possibility of incorporating feedback compensation in addition to the existing cascade compensation. With output error minimization in mind, we come up with three schemes in this category. In the process, we anticipate the risk of instability due to feedback and handle it by means of an instability preventer with an inbuilt instability detector. Extensive simulations with minimum and rionminimum phase unknown plants employing the various schemes proposed are presented. A systematic study of simulation results reveals a dyad of hierarchies of progressively enhanced overall performance. One is in the sequence of the proposed schemes and the other in going for matching more and more moments. Based on our experiments we pick one of the feedback schemes as the best. Chapter 4 is conceived of as a bridge between SISO and multivariable systems. A transition from SISO to Multi-Input Multi-Output (MIMO) adaptive control is not a proposition confined to the mathematics of dimension-enhancement. A descent from the MIMO to the SISO case is expected to be relatively simple, though. So to transit as smoothly and gracefully as possible, some issues have to be placed in perspective before exploring multivariable systems. We succinctly debate on the efforts in pursuit of the exact vis-a-vis the accurate, and their implications. We then set some notations and formulate certain results which serve to unify and simplify the development in the subsequent three chapters. We list a few standard results from matrix theory which are to be of frequent use in handling multivariable systems. We derive control laws for Single-Input Multi-Output (SIMO) systems in Chapter 5. Expectedly, SIMO systems display traits of observability and uncontrollability. Results of illustrative simulations are furnished. In Chapter 6, we formulate control laws for Multi-Input Single-Output (MISO) systems. Characteristics of unobservability and controllability stand out there. We present case studies. Before actually setting foot onto MIMO systems, we venture to conjecture on what to expect there. We work out all the cascade and feedback adaptive schemes for square and nonsquare MIMO systems in Chapter 7. We show that MIMO laws when projected to MISO, SIMO and SISO cases agree with the corresponding laws in the respective cases. Thus the generality of our treatment of MIMO systems over other multivariable and scalar systems is established. We report simulations of instances depicting satisfactory performance and highlight the limitations of the schemes in tackling the family of plants of unknown upper bound on the order and possibly nonminimum phase. This forms the culmination of our exercise which took off from the reported work involving SISO systems . Up to the end of the 7th chapter, we are in pursuit of solutions for the problem as general as in §1.4. For SISO systems, with input restrictions, the problem has been addressed in . The laws proposed there carry out adaptation only at certain discrete instants; with respect to a suitably chosen cost, the final laws are proved to be e>optimal. In Chapter 8, aided by initial suboptimal control laws, we finally devise two algorithms with continuous-time adaptation and prove their optimality. Simulations with minimum and nonminimum phase plants reveal the effectiveness of the various laws, besides throwing light on the bootstrapping and auto-rectifying features of the algorithms. In the tail-piece, we summarize the work and wind up matters reserved for later deliberation. As we critically review the present work, we decant the take-home message. A short note on applications followed by some loud thinking as a spin-off of this report will take us to finis.