 |
Distinguished Lecturers Program
Program Description
The primary purpose is to help Society chapters provide interesting and informative programs for the membership, but the Distinguished Lecture Series may also be of interest to industry, universities, and other parties. The Control Systems Society has agreed to a cost sharing plan which may be used by IEEE Chapters, sections, subsections, and student groups. IEEE student groups are especially encouraged to make use of this opportunity to have excellent speakers at moderate cost. At the request of a Society Chapter, (or other IEEE groups as mentioned above), a lecture will be scheduled at a place and time that is mutually agreeable to both the Chapter and the Distinguished Lecturer. Eighty percent (80%) of the funds for the normal travel expenses for a lecture will be paid by the Society, the remaining travel expenses will be provided by the chapter. Lecturers will receive no honorarium. Note that the group organising the lecture must have some IEEE affiliation and the lecture must be free to attend by IEEE members. The society will provide 80% of the expenses for qualified users of the program up to a maximum limit of $1000 for within a continent visit to be paid by the society and $2000 if the trip travels internationally. The speakers are geographically distributed so that this limit should be adequate for the trips to any part of the world. Travel outside of North America is encouraged provided that the society is not expected to spend in excess of $2000.
Procedures
Society Coordinator
Chairman of Distinguished Lecturers Committee School of Electrical Engineering University of New South Wales Australian Defence Force Academy Canberra ACT 2600 AUSTRALIA Ph +61 26 2688446 FAX +61 26 2688443 irp@routh.ee.adfa.edu.au
List of Lecturers
Professor Andrew Alleyne Control of Systems in a Dimensionless Framework with Applications to Vehicles The use of dimensional analysis is prevalent in several fields of the physical and life sciences. In Mechanical Engineering, the common concepts of Fluid Mechanics, Heat Transfer, and Thermodynamics are all represented by dimensionless variables. These include the well known dimensionless numbers such as the Reynolds Number, Froude Number, Nusselt Number, etc. The question we raise and hope to answer in this talk is whether or not the field of Systems and Control can benefit from dimensionless analysis as other fields have done. This talk will begin by detailing the initiation of our study into dimensionless analysis. A scaled testbed for vehicle dynamics and control studies has been developed at the University of Illinois. In order to validate controller designs for scaled vehicles and ensure their validity on full sized vehicles it was important to guarantee dynamic similitude between the scaled and full sized dynamical systems. We detail the derivation of a dimensionless form for the linear dynamics of vehicles and demonstrate how we ensure our scaled vehicles are dynamically similar to full sized ones. It can be shown that the dimensionless form for this linear system can be thought of as a very convenient transformation of the original dimensional system. Subsequent to presenting a dimensionless form for the vehicle dynamics, we illustrate several key benefits that we have found from working in a dimensionless framework. First, it is possible to uncover underlying dynamical relationships that do not seem clear when studying the dimensional system dynamics. Secondly, the parametric uncertainty associated with nominal vehicle representations is greatly reduced in a dimensionless framework, thereby leading to less conservative controller constraints. Finally, parametric interdependence is more easily uncovered and can be used to greatly reduce system excitation requirements for identification or adaptation mechanisms. Precision Control for Micro-Scale Manufacturing Established tools for the manufacture of systems with micro-scale features include standard lithography techniques. This talk examines a different method for microscale manufacture called Microscale Robotic Deposition (MRD). This is a solid freeform fabrication technique where by material is deposited to rapidly "build up" complex 3 dimensional structures. Current systems under construction focus on the creation of periodic, lattice-like structures for their potential of acting as waveguides. The talk centers on the control aspects of the MRD problem, particularly the precision motion control problem. The periodic nature of the systems under construction allow for the use of an Iterative Learning Control (ILC) approach to positioning problem. Iterative Learning Control utilizes information from a previous attempt at a task to update a feedforward signal and better the tracking performance on future attempts at the identical task. Most practical ILC designs incorporate some stabilizing filtering mechanisms to provide robustness to the learning process. In this talk we will extend that approach be introducing an adaptive filtering technique based on Time-Frequency analysis. Controls and Experiments: Lessons Learned This talk describes different lessons that can be learned by including experimental aspects into control system research. Several key lessons are identified and then each lesson is developed within the context of a particular experimental system. A variety of physical experimental systems are used to illustrate that the key lessons need not all be found in the same system but, should one be working with a variety of systems, it is likely that one or more of these issues would arise. For the purpose of the IEEE Controls Systems Society Distinguished Lecturer Program, the particular experimental systems used in this talk can be chosen to match the interest areas of the particular audience. The actual experimental systems available span the fields of Vehicle Dynamics, Manufacturing, Aerospace Systems, Air Conditioning and Refrigeration, and Fluid Power. However, it is felt that the main points of this talk can easily be applied to many other fields.
Professor Anu Annaswamy Active-adaptive Control of Acoustic Resonances in Flows Several fluid flow problems related to propulsion and power generation exhibit strong acoustic resonances. Produced due to interactions of the acoustics with other underlying unsteady mechanisms such as unsteady heat-release or shear flow instability, these resonances manifest as large and sustained pressure oscillations. In addition to the obvious undesirable effect of high ambient noise and acoustic fatigue, these oscillations are also coupled with other damaging effects such as excessive vibrations, high burn rates, lift-loss, and ground erosion. Compromises in reducing these oscillations lead to departures from the desired operating conditions and can lead to suboptimal performances with reduced heat-output, increased emissions, or decreased efficiency. Over the past few years, active control technology has been increasingly sought after to realize the desired performance metrics in these problems without encountering resonant behavior. In order to provide guaranteed and uniform performance over a large range of operating conditions in the presence of various system uncertainties, it has been demonstrated in these problems that a model-based approach to designing the control strategy is feasible and scalable, and leads to a reliable and improved pressure reduction at the desired operating conditions. In this talk, two examples of such fluid flow problems, combustion-instability and impingement-tones in supersonic flows, and their active control will be discussed. Linear and nonlinear models of the resonant mechanisms based on both physically-based and system-identification principles will be presented. In active-adaptive control of combustion systems, linear strategies based on H_2 and H_infinity control, open-loop methods based on slow switching, and adaptive posicast control methods, and their impact on a range of rigs from a 1 kW benchtop combustor to a 4 MW engine-replica test-rig will be discussed. In active-adaptive control of supersonic impingement tones, a POD-based active control strategy and the corresponding experimental results from a STOVL supersonic jet facility at Mach 1.5 will be presented.
Professor Karl Johan Åström Control - The Hidden Technology The field of automatic control is about 50 years old. This paper presents some reflections on the dynamic development of the field. It starts with a brief history and a discussion of engineering science and natural science. Automatic control being the first systems field was a paradigm shift because it fitted poorly in organizations based on mechanical, electrical and chemical engineering. Key ideas in the development of the field are presented. The interplay of theory and applications are discussed as are relations to other fields such as mathematics and computer science are discussed. It is attempted to make an assessment of the current status. The lecture ends with some speculations about future development. Questions dealing with research, education and applications will be covered. An explanation of the title of the paper will also be given. Friction Models and Friction Compensation Friction appears in all mechanical systems and has a significant impact on control. Successful design of mechatronic systems requires an understanding of the effects of friction as well as techniques to compensation. Friction phenomena are complicated because they are caused by many different physical mechanisms. Friction can cause a substantial deterioration of the performance of a control system. Typical effects are steady state errors and oscillations. Many attempts have been made to compensate for friction. Early efforts include introduction of dither signals. Other ideas are to based on model based control. This paper reviews several models for friction that have been useful to model friction in motion control systems. The models discussed include Dahl's model, the Bliman-Sorine model and the LuGre model. Different techniques for friction compensation are also discussed and results from practical experiments with friction compensation are presented. Modeling of Physical Systems Modeling and simulation have experienced an amazing development since its beginning in the 1920s. At that time, the technology was available only at a handful of university groups. Today it is available on the desk of all engineer who needs it. The paper presents the current status of modeling and simulation. It draws on the historical perspective to explain how the field has developed. Particular emphasis is given to shifts in technology and paradigms. Even if the technology of modeling has advanced considerably the standard tools used today are very similar to the ones used 40 years ago. Recently there has been several attempts to develop methodologies that are much more appropriate for modeling of complex technical systems. These methodologies draw from object oriented methodology in computer science, differential algebraic equations in numerical mathematics and control theory. The tools Dymola, Omola, Omsim and Modelica which all have their origins in Lund are discussed. An example of modeling of a thermal boiler is used for illustration. Tuning and Adaptation of PID Controllers Feedback is a very powerful idea. Its use has often had revolutionary consequences with drastic improvements in performance. Credit is often given to a particular form of feedback although it frequently is feedback itself that gives the real benefits and the particular form of feedback used is largely irrelevant. The PID controller is by far the most dominating form of feedback in practical use today. More than 90% of all control loops are based on PI or PID control. Applications cover a wide range process control, motor drives, magnetic and optic memories, automotive, flight control, instrumentation, etc. PID controllers have been augmented substantially by facilities for feedforward, automatic tuning, gain scheduling, adaptation and diagnostics. The lecture gives a broad presentation of the development of PID control with particular emphasis on recent developments in use of feedforward, compensation for time delay, tuning and adaptation.
Dr Michèle Basseville Early Warning of Small Deviations for Condition-based Maintenance: Theory and Application An increasing interest in condition-based maintenance has appeared in a large number of industrial applications. The key idea is to replace regular systematic inspections with condition-based inspections, in order to prevent from a possible malfunction or damage before it happens. Condition-based inspections are decided upon the continuous monitoring of the considered system (machine, structure, process or plant), based on the available measurements. For achieving this, without affecting the process safety, availability and performances, it is of crucial importance to be able to perform the early detection and isolation of slight deviations of the process with respect to a reference behavior considered as normal. It turns out that there exists a mathematical theory providing us with tools which perform the early detection task. This theory is known under the name of the statistical local approach, and it is especially suited to component faults. It aims at transforming a large class of detection problems concerning a parameterized stochastic process into the universal problem of monitoring the mean of a Gaussian vector. The isolation task is then performed by inspecting the components of this Gaussian vector, based on standard techniques for dealing with nuisance variables. The purpose of the lecture is to describe both the key components of this theory and real industrial examples which have been successfully addressed by applying its principles. Strong emphasis will be put on the key issues in the use of this approach for component faults in linear and nonlinear dynamical systems. This theory assumes that a model of the monitored system is available. This is a reasonable assumption, since many industrial processes rely on physical principles, which write in terms of (differential) equations, providing us with (dynamical) models. The use of (physical) parameters is mandatory when isolation and diagnosis are required. But when physical models are either too complex or not known at all, semi-physical or black-box models (neural networks, wavelet networks) can be used instead, in the framework of the local approach, enlarging its initial scope. Output-only Structural Vibration Monitoring for Damage Detection and Localization In-operation vibration monitoring for complex mechanical structures and rotating machines is of key importance in many industrial areas. Typical examples are offshore structures subject to swell, buildings subject to wind or earthquake, bridges, dams, wings subject to flutter in flight, and turbines subject to steam turbulence, friction in bearings, and imperfect balancing. These systems are subject to both fast and unmeasured variations in their environment and small slow variations in their vibratory characteristics. Of course, only the latter are to be detected, using the available data (e.g. accelerometers), and noting that the changes of interest (1% in eigen-frequencies) are not visible on spectra. The key issues are to identify and monitor vibrating characteristics (modes and modal shapes) of mechanical structures and machines subject to unmeasured and non-stationary natural excitation. The purpose of the lecture is to describe the key components of and real experimental results obtained with a set of algorithms performing modal analysis and vibration monitoring. These algorithms process multi-sensors measurements recorded during routine operation, without artificial excitation, nor slow down, nor shut down of the machine, and possibly by different sensor pools. The approach is based on the modeling of modes and modal shapes as eigenvalues and observed components of the eigenvectors of the state transition matrix of a linear dynamical system; on the use of output-only and covariance-driven identification methods (such as instrumental variables, balanced realization, and subspace-based algorithms); and on the computation of specific chi^2-type tests for fatigue or damage detection. These tests evaluate the importance, for each mode, of the modal deviation which is detected. They can be equally used for laboratory test beds with measured excitation, and for in-operation data without measuring the (possibly non-stationary) excitation. For diagnosing (in terms of modifications of e.g. volumic mass or Young modulus) and localizing the fatigues or damages, the proposed solution circumvents the resolution of the ill-posed inverse problem which results from the use of a (generally) high dimensional and not identifiable physical F.E. model. The approach relies upon the projection and aggregation of the elementary changes in the physical space onto the (much lower dimensional) space of identified odes and modeshapes, and upon the computation of convenient Jacobian matrices fed into the above chi^2-type test. These monitoring and diagnostics algorithms have been generalized to models more complex than those (linear) for vibrations, and can be used for monitoring in industrial process control (gas turbine, electric or thermal plant) or for on-board diagnostics (catalytic exhaust).
Professor Robert R. Bitmead Cautious Controller Tuning The problem considered is the use of closed-loop experimental data to guide controller adjustment or tuning. The main problem is to move the controller from one stabilizing value to another with a stability and/or performance guarantee. The chief tool is the $\nu$-metric of Vinnicombe, which links the allowable $\nu$-distance between controllers and the current closed-loop performance. A number of scenarios are considered which treat cases in which different information is available about the closed-loop system. This includes estimation of the stability margin from closed-loop data and joint model and controller adjustment. This seminar has a strong theoretical basis with links into the algebra of stabilizing controllers and the analysis of metric spaces. Nevertheless, the nature of the results is very informative of practical issues linked with the safe tuning of controllers (including restricted complexity controllers) using closed-loop measurements. State Estimation in Model Predictive Control Model Predictive Control (MPC) is a currently popular and applicable control method in which one solves on-line an explicit constrained optimal control problem and applies part of the solution. The attractions are that constraints may be included on states, inputs and outputs, that common optimization software may be used, and that it uses significant computational resources in the control loop which make it applicable to many multi-input/multi-output and nonlinear systems. One of the drawbacks of MPC is that it is posed as a full state-feedback control, when in practice there is a need to employ state observers to compute a best-possible but inaccurate state estimate. The key problem with including state estimates is the need to accommodate the satisfaction of the constraints. In this presentation, an approach will be presented in which the covariance of state estimate - a quantity frequently provided by observers but then discarded - might be used to amend the constraints in a straightforward fashion. This then will be linked to coordination between systems and the design of information architectures for control. A Skeptical Introduction to Combustion Instability Modeling Combustion is the central process of jet engines and gas turbines. Pollution can be decreased and efficiency increased if these machines can be operated with leaner mixtures - lower fuel-to-air ratios - leading to cooler and more complete combustion. However, as the fuel ratio is decreased, the flame front begins to become unstable and demonstrates a periodic limit cycle. The variations introduced by this cycling more than cancel the benefits of the leaner mixture. So we are looking for feedback control solutions to stabilize the flame. This seminar is not about this. It is about modeling for control to help achieve this. After presenting the obligatory cut-away pictures of jet engines, the presentation will deal with the travails of fitting to experimental limit-cycling data a linearized model suitable with the aim of using the resultant model for control. The problem is that such periodic data really do not contain much information. Nevertheless, a model has been fitted which appears capable of replicating the data. This will be developed using describing function ideas. Next a validation of the identified model (focused on likely utility for control design) needs to be performed. The construction of the validation experiment will be described, invoking excitation of coupled nonlinear attractors. The problem now would appear to be the interpretation of these validation experiment results - several conundra arise. The seminar will deal with issues of nonlinear modeling from data using physical model structures. It should appeal to control and signal processing researchers, as well as those interested in impressing people by saying they study Rocket Science. Carpentry and Mechanics of Joint Identification and Control: data preparation and iterative control of a sugar cane crushing mill We deal with the "nuts and bolts" aspects of iterative system identification and control design on a specific process of a sugar cane crushing mill from North Queensland, Australia. Our emphasis is on the aspects in which the process knowledge is used to guide the myriad design choices necessary before resorting to the use of special-purpose software for identification or control design. Of particular concern is the data selection and prefiltering which should precede system identification and which has the capacity to connect the a priori knowledge, the design objective and the computational model fitting exercise.
Professor Stephen Boyd Advances in Convex Optimization: Interior-point Methods, Cone Programming, and Applications In this talk I will give an overview of some major developments in convex optimization that have emerged over the last ten years or so. The basic idea is that convex problems are fundamentally tractable, in theory and in practice. The polynomial worst-case complexity results of linear programming have been extended to nonlinear convex optimization, and interior-point methods for nonlinear convex optimization achieve efficiencies approaching that of modern linear programming solvers. Moreover, special purpose implementations for large-scale applications can take advantage of many different types of problem structure. Several new classes of standard convex optimization problems have emerged, including semidefinite programming and linear matrix inequalities, which are well known in control, and also determinant maximization, second-order cone programming, and geometric programming, which are less well known. Like linear and quadratic programming, we have a fairly complete duality theory, and very effective numerical methods for these problem classes. There has been a steadily expanding list of new applications of convex optimization, in areas such as circuit design, signal processing, statistics, communications, and other fields. Convex optimization is also emerging as an important tool for hard, non-convex problems. Convex relaxations of hard problems provide a general approach for handling hard optimization problems, with applications in combinatorial optimization, analysis and design of nonlinear and uncertain systems, and robust optimization. (Joint work with Lieven Vandenberghe) Optimal Design of Analog CMOS Circuits via Geometric Programming As more ICs incorporate some analog functionality, the demand for analog CMOS design is greatly increasing. Not only will the number of new designs increase, but so will the number of designs that must be ported every 12 months or so to the newest process technology. This increase in demand comes during a great shortage of experienced analog designers. Predictions abound that analog design will become a critical bottleneck for next-generation system-on-a-chip (SoC) designs. In this talk we describe a new general method for optimized design of analog CMOS circuit blocks. We first cast the circuit design problem as a very special type of mathematical optimization problem, called a geometric program, which we then solve using recently developed interior-point methods. The result is a method that can very efficiently determine globally optimal circuit designs, given from specifications (such as on power, area, open-loop gain, bandwidth, etc.). The method can be used for extremely rapid design, performance and trade-off analysis, and porting of analog circuit cells to new technologies. After giving some general background, we describe the underlying optimization method, and explain how the method applies to the design of a common op-amp. (Joint work with Mar Hershenson and Tom Lee) Fastest Mixing Markov Chain on a Graph We consider an undirected graph, with edges labeled with the probability of a transition between the associated vertices. The associated Markov chain has a uniform equilibrium distribution; the rate of convergence to this disitribution is determined by the second largest (in magintude) eigenvalue of the transition matrix. We show that determining the fastest mixing Markov chain, with a fixed graph, is a convex problem that can be expressed as an SDP. (Joint work with Persi Diaconis and Lin Xiao)
Professor Marco Campi The Scenario Approach to Robust Control: Randomized Solutions and Confidence Levels Many worst-case robust control problems cannot be solved due to computational difficulties. In this talk, a new probabilistic solution framework is proposed for robust control analysis and synthesis problems that can be expressed in the form of robust convex optimization. This includes for instance the wide class of NP-hard control problems representable by means of parameter-dependent linear matrix inequalities (LMIs). By appropriate sampling of the constraints, one obtains a standard convex optimization problem (the scenario problem) whose solution is approximately feasible for the original (usually infinite) set of constraints, i.e. the measure of the set of original constraints that are violated by the scenario solution rapidly decreases to zero as the number of samples is increased. Explicit and efficient bounds on the number of samples required to attain a-priori specified levels of probabilistic guarantee of robustness are given. A rich family of control synthesis problems which are in general hard to solve in a deterministically robust sense is therefore amenable to polynomial-time solution, if robustness is intended in the proposed risk-adjusted sense. The approach is illustrated by means of significant examples in robust control. Tuning Industrial Controllers: The Virtual Reference Feedback Tuning (VRFT) Approach In many practical control applications, a mathematical description of the plant is not available and the controller has to be designed on the basis of measurements (data-based design.) This problem has attracted the attention of control engineers since the forties with the pioneering work of Ziegler and Nichols. In this talk, a new approach to data-based controller design named Virtual Reference Feedback Tuning (VRFT) is presented. The main features of VRFT are: Decision Making in an Uncertain Environment: the Scenario-Based Optimization Approach.
A central issue arising in many engineering problems is to make a decision in spite
of an uncertain environment. Application examples are:
In this talk, we present a computationally efficient methodology for dealing with
uncertainty in optimization and decision making based on sampling a finite number of
instances (or scenarios) of the uncertainty. The following points are discussed:
Professor Christos G. Cassandras An Introduction to Discrete Event Systems and their Applications Over the past few decades, the rapid evolution of computing, communication, and sensor technologies has brought about the proliferation of new dynamic systems, mostly technological and often highly complex. Examples include: computer and communication networks; automated manufacturing; air traffic control; highly integrated command, control, communication, and information systems; intelligent transportation; distributed software; and so forth. Almost all of the activity in these systems is governed by operational rules designed by humans; their dynamics are therefore characterized by asynchronous occurrences of discrete events, some controlled (like hitting a keyboard key) and some not (like a spontaneous equipment failure or a packet loss), some observed by sensors and some not. These features lend themselves to the term "discrete event system" for this class of dynamic systems. This talk will describe the fundamental differences between discrete event (or, event-driven) systems and conventional time-driven systems and discuss appropriate modeling frameworks for the former, such as automata and Petri nets. It will then present some representative control and optimization problems as they manifest themselves in discrete event systems (drawn from "real world" experience with manufacturing systems, computer networks, elevator dispatching control, and air traffic control) and overview solution approaches, both analytical and based on computer simulation methods. An Introduction to Hybrid Systems and Some Applications to Manufacturing System Integration Complex dynamic systems are often decomposed into a lower-level component with time-driven dynamics (describing the physical state of the system) interacting with a higher-level component with event-driven dynamics (describing the changes in the system operating modes). This gives rise to a Hybrid System. In order to analyze and control such a system, one must combine knowledge from classical control theory, based on differential/difference equations, with more recent developments in discrete event systems. This talk is intended to describe some modeling frameworks for hybrid systems and to present examples ranging from simple switching to complicated hierarchical systems. A natural setting of this type is encountered in manufacturing processes: The physical characteristics of production parts undergo changes at various operations described by time-driven models (such as differential equations), while the timing control of operations is described by event-driven models (such as timed automata). Accordingly, manufactured parts are characterized by physical states (e.g., temperature, geometry) subject to time-driven dynamics, and by temporal states (e.g., operation start and stop times) subject to event-driven dynamics. The tradeoff between the physical "quality'' of parts and various timing requirements on part delivery leads to optimal control problems which are typically nondifferentiable and nonconvex. The talk will present at an intuitive level how these difficulties can be overcome and how simple explicit solution methodologies can be obtained, which will be illustrated for common manufacturing processes through the use of interactive java applets. Joys and Perils of Automation One of the definitions of the word "control" is "to govern or direct according to rule" (Merriam-Webster dictionary). In science and engineering, these "rules" have traditionally been dictated by the laws of nature - such as gravity or conservation of mass. Computer technology, however, has enabled us to build complex systems that have become essential to our daily life, from automated factories to computer networks, with intelligent highways and pilotless planes not too far in the horizon. The "rules" that these systems must obey are as arbitrary as human imagination can make them (as in designing a video game where one may create a virtual world where anything goes). While this is exciting, it is also dangerous - it takes but one minor "bug" or "virus" to bring a multimillion factory to a standstill, the Internet to crash, or the Mars exploration vehicle to erroneously "think" that its landing legs were deployed, effectively forcing it to commit electronic suicide. Many of the dangers of automation stem from the lack of designers and engineers with appropriate skills that are cultivated through an understanding of what a "system" is and how to evaluate the effectiveness of a controller before deployment. This presentation will employ simple computer simulation examples to illustrate the difference between physical processes subject to the laws of nature and human-made processes that must satisfy human-made rules. We will then show how "automatic control" can be used and demonstrate both its benefits and risks.
Professor Michel Gevers Iterative Feedback Tuning: Theory, Applications and Extensions One of the most active areas of research in the nineties has been the study of the interplay between system identification and control design. This study has led to the development of "iterative identification and control design schemes". The analysis of these iterative schemes, and the observation that the model in these iterative schemes is only a tool for the computation of the controller parameters, has in turn led to the development of a direct iterative controller tuning scheme, in which the controller parameters are estimated by a direct optimization of a control performance criterion. The main difficulty in such direct optimization had always been that the gradient of the performance criterion is a function of the unknown system, hence the need for a model estimate. A breakthrough was achieved in 1994 when it was shown that an unbiased estimate of this gradient can be obtained from an experiment performed on the unknown system, without the need for a model. This led to an iterative minimization of the performance criterion, in which the gradient is performed at each iteration by performing this special experiment on the present closed-loop system. This special experiment consists of feeding back the output of the system to the reference input; hence the name Iterative Feedback Tuning (IFT) given to this model-free controller tuning scheme. Since its inception in 1994, IFT has been applied to a large number of industrial and mechanical applications. It has been extensively analyzed, and special variants have been developed for specific applications and objectives. In this presentation, we will present the very simple theory of IFT, survey some of the applications, and present some extensions, in particular a variant aimed at the tuning of PID controllers with minimum settling time. Ten Years of Progress in Identification for Control "Adaptive control" and "dual control" are two early instances in which parameter estimation and control design were formulated as a joint optimization problem. However, dual control has never been applied to realistic problems because of the curse of dimensionality: the computational load is just enormous, even for simple problems. As for adaptive control, most of the results have focused on the situation where a full-order model was used. The study of identification and control as a combined design problem, in the context of restricted complexity model structures (i.e. with unmodelled dynamics) began in earnest around 1990, when the interplay between identification design and control design became a topic of intense research. The first few years of work focused on the manipulation of the bias error distribution: how to obtain a nominal model that is suitable for the design of a controller, which ought to achieve good performance on the real system. It soon became clear that the nominal and the achieved closed loop system must be close to one another, where closeness is measured in a norm that is a function of the control performance criterion. Given that these two closed loop systems are a function of the `to be designed controller', this requirement led to iterative schemes, based on a sequence of model updates and controller updates. In the late nineties, attention turned to the manipulation of the variance error distribution, i.e. how should one tune the identification experiment in such a way that the model uncertainty set is suitable for the design of a robust controller? Necessary and sufficient conditions have been obtained for a controller to stabilize all models in an uncertainty set obtained by Prediction Error identification methods, and connections have been established between model uncertainty sets and corresponding sets of stabilizing controllers. This talk will present a tutorial overview of these successive advances in identification for control over the last decade. The iterative identification and controller design schemes that have emerged from this area of research activity have found their way very quickly into industrial applications, thus showing that this area is not only full of challenging theoretical problems, but that it is also of high interest for the practitioners.
Professor Tohru Katayama Subspace System Identification: An Introduction Subspace methods have attracted much interest for more than a decade. They involve geometric operations on subspaces spanned by the column or row vectors of certain block Hankel matrices formed by the input-output data. These operations are best performed by the SVD and QR decomposition. Advantage of subspace methods is that the problem of local minima and the difficult model selection problem inherent in the classical prediction error approach to MIMO identification are avoided, except for the estimation of the dimension of the state space. This talk gives an elementary introduction to the subspace identification method from the point of realization theory. First a brief review of the classical prediction error methods is given. Then, I would like to stress the importance of the realization theory to better understand the various forms of subspace identification algorithms (MOESP and N4SID), thereby giving some basic knowledge to read advanced papers. Some numerical results are also included. This talk is kept elementary so that many people understand the present status of subspace identification methods.
Professor Jan Maciejowski An Overview of Robust and Multivariable Control This lecture gives an overview of the main developments in modern robust and multivariable control theory. It assumes that the audience has some knowledge of "classical" methods and ideas of feedback design (Bode plots, stability margins, sensitivity function, etc). The lecture makes some use of matrices and of state-space models, but the emphasis is on explaining the ideas rather than on the technical details. The lecture covers the following topics: Current issues in Model Predictive Control: Efficiency, Feasibility, Priorities This lecture gives a personal view of certain "hot topics" in Model Predictive Control (MPC). MPC is the only "Advanced Control" methodology to have gained wide acceptance in (certain sectors of) industry. The fundamental theoretical advantage of MPC over other control methodologies is its ability to take constraints into account in a systematic manner. In my opinion this is actually more important in most applications than the fact that it optimises a cost function. Although MPC was developed largely in the petrochemical sector, and is now being applied in other process control sectors, I am sure that it will also be found of great benefit in many other "non-process" applications. (Very many control problems can be posed as "operate as close to the constraints as you can".) As the range of applications of MPC grows, the following three aspects (at least) will become the "bottleneck" questions: My talk will outline the main ideas and developments - and they are considerable - which are found in the research literature, addressing these key questions. The views expressed in this lecture were developed while writing my recently-completed book: "Predictive Control with Constraints", to be published by Pearson Education (Prentice-Hall) in 2001.
Professor A. Stephen Morse Supervisory Control of Families of Linear Set-Point Controllers This talk will describe a simple, "high Level" controller called a "supervisor" which is capable of switching into feedback with a siso process, a sequence of linear positioning or set-point controllers from a family F_C of candidate controllers, so as to cause the output of a process to approach and track a constant reference input. The process is assumed to be modeled by a siso linear system whose transfer function is in the union of a number of subclasses, each subclass being small enough so that one of the controllers in F_C would solve the positioning problem, were the process's transfer function to be any member of C_i. The supervisor decides which controller to put in feedback with the process, not by an exhaustive search-i.e., by experimentally evaluating each and every candidate controller's performance by briefly aplying it to the process-but rather by continuously comparing in real time suitable defined "output estimation errors" generated by the candidate controllers, whether or not they are in feedback with the process. It is shown that under reasonable mild conditions, the supervisor can successfully perform its function ins pite ofmodeling errors, provided the errors are sufficiently small. It is also shown that the supervisor will invariably correctly classify the process in finite time, so long as the reference input is nonzero and the "dc gains" of the "nominal" candiate process model transfer functions are distinct. Logic-Based Switching: A Form of Intelligent Control Recent advances in system theory have shown that much can be gained by using logic-based switching strategies, together with more familiar techniques in the synthesis of feedback controls. The overall models of systems composed of such logics together with the processes they are intented to control, are concrete examples of what might be called "hybrid dynamical systems". In this talk we wil describe three different hybrid systems of this type-each consists of a continuous-time process to be controlled, a family of continuous-time, candiate fixed-parameter or adaptive controllers, and an "event-driven switching logic". The first two logics called "hysteresis switching" and "dwell-time switching" respectively, are simply strategies capable of determining in real time which candidate controller should be put in feedback with a process in order to achieve desired closed-loop performance. The third, called "cyclic switching" has been devised to deal with the well-known certainty equivalence stabilizability problem which arises in the synthesis of identifier-based adaptive controllers because of the existence of points in parameter space where the design model upon which certainty equivalence synthesis is based, loses stabilizability.
Professor Thomas Parisini Learning approach for fault detection, isolation, and accomodation of nonlinear uncertain systems This lecture is based on recent joint research work with Prof. M. Polycarpou and Dr. X. Zhang. The objective of the lecture is to give a tutorial overview of a novel learning approach for designing and analyzing fault detection, isolation and accomodation schemes for a class of nonlinear uncertain systems. The general approach is based on a architecture made of two basic modules: Monitoring module: provides the information about the detection of a fault and the information about the specific fault that occurred in a class F of a priori-specified fault structures. This module is made of a bank of nonlinear adaptive estimators. One of the nonlinear adaptive estimators is the fault detection and approximation estimator (FDAE) used for detecting and approximating faults. An on--line approximation model, typically based on neural approximators, is used in the FDAE. The remaining ones are fault isolation estimators (FIEs) used only after a fault is detected for isolation purposes. Each FIE corresponds to a particular type of fault in class F. Under normal operating conditions (without faults), the FDAE is the only one monitoring the system. Once a fault is detected, then the bank of FIEs is activated and the FDAE goes into the mode of approximating the fault function. The case that none of the isolation estimators matches the occurred fault (to some reasonable degree) corresponds to the occurrence of a new and unknown type of fault, and the approximated fault model can then be used to update the fault class F and also the bank of isolation estimators. The fault model generated either by the isolation estimators (in the case of a match) or the detection/approximation estimator is used for fault diagnosis and provides the information to be used by the controller module for fault accommodation. Controller module: is designed to reduce the tracking error in the presence of modeling uncertainty and the possible occurrence of faults. This module exploits information coming from the monitoring module. Specifically, based on the fault information obtained during the fault diagnosis procedure, a fault-tolerant control component is designed to compensate the effects of faults. In the presence of a fault, a nominal controller guarantees the boundedness of all the system signals until the fault is detected. Then the controller is reconfigured after fault detection and after fault isolation, respectively, to improve the control performance using the fault information generated by the diagnosis module. In the lecture, a complete design and analysis of the above scheme will be presented in a tutorial but rigorous way and some simulation examples will also be reported, showing the practical aspects involved. Approximating networks for the solution of functional optimization problems This lecture is based on recent joint research work with Prof. R. Zoppoli and Dr. M. Sanguineti. The objective of the lecture is to give a tutorial overview of an approximate approach to solve functional optimization problems that typically arise in control and estimation contexts. Functional optimization problems can be solved analytically only if special assumptions are verified (typically, if we handle dynamic systems, they have to be linear; cost functions have to be quadratic; random variables, if present, have to be Gaussian). Otherwise, approximations are needed. The approximate method we propose is based on two steps. 1) The decision functions are constrained to take on the structure of linear combinations of basis functions containing ``free'' parameters to be optimized (hence this step can be considered as an extension to the Ritz method, for which ``fixed'' basis functions are used). Then the functional optimization problem can be approximated by a nonlinear programming one. Linear combinations of basis functions are called approximating networks when they benefit by suitable density properties. We term such networks nonlinear (linear) approximating networks if their basis functions contain (do not contain) free parameters. Several classes of neural approximation schemes can be used be used in this respect and in the lecture some neural scheme will be described in some detail. For certain classes of d-variable functions to be approximated, nonlinear approximating networks may require a number of parameters increasing moderately with d, whereas linear ones may be ruled out by the curse of dimensionality. Since the cost functions of the resulting nonlinear programming problems include complex averaging operations, we minimize such functions by stochastic approximation algorithms. As important special cases, we address stochastic optimal control and estimation problems. Numerical examples show the effectiveness of the method in solving optimization problems stated in high-dimensional settings, involving, for instance, several tens of state variables.
Professor Kevin Passino Systems Biology of Swarm Cognition Systems biology of decision making focuses on understanding the structures, dynamics, and evolution of complex interconnected biological mechanisms that support decision making by individuals and social animal groups. In this talk, an experimentally validated mathematical model of the nest-site selection process of honey bee swarms is introduced. In this spatially distributed dynamical feedback process individual bee actions and bee-to-bee communications combine to produce an emergent "consensus" nest choice. Using a key idea of "group memory", the process has connections to neurobiological cognition systems, especially at the behavioral level: the swarm can effectively discriminate between different quality nest sites and eliminate from consideration relatively inferior distractor sites. Simulations indicate that individual-level bee decision-making mechanisms have been tuned by natural selection to provide a balance between the need for fast and accurate decisions at the group level. For more information see: http://www.ece.osu.edu/~passino/ Stable Cooperative Agent Distributions: Theory and Applications in Biology and Engineering The problem of how to perform decentralized coordination of the spatial distribution of agents is found in a number of applications. Autonomous robots with range-limited communication and sensing radii must coordinate their decisions on where to allocate their sensing resources to achieve prioritized surveillance of multiple widely dispersed areas. A hive of honey bees allocates foragers according to forage site profitability without a central coordinator. In this talk, a mathematical model of the behavior of a group of agents and their interactions in a shared environment is introduced. Environmental spatial constraints are included to model range-limited sensing, motion, and communication capabilities of the agents. General sensing, coordination, and motion conditions on the agents are derived that guarantee that an "ideal free distribution" (IFD) of the group of agents will emerge. The impact of group size and environment type on the distribution of agents is quantified. Applications of the theory to a multivehicle cooperative surveillance problem and honey bee social foraging are discussed. For more information see: http://www.ece.osu.edu/~passino/
Professor Anders Rantzer On Convexity and Relaxation in Nonlinear and Hybrid Control Since the 1950's, the idea of dynamic programming has propagated into a vast variety of applications. The basic Hamilton-Jacobi-Bellman equation is general and very simple, but the "curse of dimensionality" is often prohibitive and restricts most applications to a discrete state space of moderate size. In the last few years, there has been several exciting developments related to convexity of the corresponding Bellman inequality, both theoretical and computational. For example, research on continuous nonlinear control has generated analogs of the primal-dual optimization used in discrete shortest path algorithms. In nonlinear control, this has resulted in new stability criteria and new methods for control synthesis by convex optimization. Another very exciting development is based on approximations of the cost to go. It turns out that the computational complexity of traditional dynamic programming algorithms often can be drastically reduced by relaxing the demand for optimality in the Bellman inequality. In fact, finding a solution which is guaranteed to be within 10% from the optimum turns out to be drastically less expensive than finding one within 1%. Results of this type will be discussed and illustrated by examples. Density and Flow: A Different View on Nonlinear Control The concept of "denisty functions" has recently been introduced as a tool for stability analysis of nonlinar ordinary differential equations. Statements are given in terms of "almost all trajectories" of the system. The theory is similar to the classical theory of Lyapunov, but the implications are weaker and applicable to many situations where global stability in the classical sense does not hold. Density functions also have remarkably nice properties for control synthesis. While the set of control Lyapunov functions for a given system may not even be connected, the corresponding set for density functions is always convex. This opens new possibilities for nonlinear control synthesis by convex optimization. In this seminar we discuss the duality between density functions and Lyapunov functions and a converse theorem that proves existence of a density function from assumptions on almost global stability. Simple examples of pendulum dynamics turn out to be illuminating.
Professor Karen Rudie Minimal Communication in a Distributed Discrete-Event Control System Distributed discrete-event systems, in which agents (or local sites) are required to communicate in order to perform some specified monitoring and control tasks, are considered. Each agent is modeled as a finite-state machine that must be able to distinguish between its states to perform some required task. To help it disambiguate states, an agent uses a combination of direct observation (obtained from sensor readings available to that agent) and communicated information (obtained from sensor readings available to another agent). Since communication may be costly, a strategy to minimize communication between sites is developed. The complexity of the solution reflects the interdependence of the agents' communication protocols. That is, the decision to communicate an event relies on which event sequences are indistinguishable to an agent, which, in turn, is a result of what has already been communicated to that agent. Knowledge is a Terrible Thing to Waste: Using Formal Reasoning about Knowledge and Inference to Solve Discrete-Event Control Problems Discrete-event systems are processes whose behaviour can be characterized by sequences of events and can be represented by finite-state automata or directed graphs. Control problems arise because the systems can generate undesirable sequences. Work in this area typically addresses when it is possible to derive agents that can prohibit bad sequences. These problems are more difficult computationally if they must be solved using decentralized control, where each agent has only a partial view of overall system behaviour. A mathematical theory of knowledge, as developed by Halpern & Moses, has been used to model distributed systems. This theory provides a formal way to reason about what processes "know". The model is based on a modal logic and uses Kripke structures, which provide visual pictures of what each agent "knows" and "does not know" Using this formalism, we model decentralized discrete-event problems. We show that problem solution amounts to determining whether for each required action, at least one agent "knows" what event to disable. When a supervisor cannot make a definitive control decision based on its own knowledge of the system, the supervisor may reason about whether other supervisors have sufficient knowledge to eventually make the correct control decision.
Dr. Hebertt Sira-Ramirez Sliding Mode Control Design Controller design is approached from a combination of passivity based control, sliding mode control and differential flatness. A tutorial introduction is provided to these three powerful nonlinear controller design techniques with numerous simple and motivating examples ranging from power electronics to chemical process control. The advantageous combination of the three techniques provide 1) simple, 2)robust and 3) natural nonlinear feedback controllers in the following senses: 1) Simplicity is obtained from the passivity based control approach. The passivity based controller design does not destroy the useful nonlinearities represented by the dissipative structure of the nonlinear system. The approach is not constrained to Lagrangian systems but it is applicable to any affine nonlinear system representable by sufficiently differentiable vector fields. We propose a quite natural drift vector field with respect to constant level sets of the energy storage funcion of the system. The decomposition results in a simple, yet design oriented identification of the dissipative components, the invariant components and the non dissipative components. This decomposition is based on a given energy storage function of the system, but useful storage functions can also be proposed following various other techniques available in the literature. Such decomposition is inspired in classical "force" decomposition theorems available from analytical mechanics. A general justification of the energy shapping plus damping injection controller design method is also furnished from simple considerations about the given system fields. 2) The Robustness of the approach is inherited from the sliding mode control approach. An exogenous system, normally used for the nonlinear dynamical feedback controller design in the passivity based method, is forced to follow a stabilizing trajectroy, via sliding mode control. The exogenous system shares the same control as the plant. As a consequence, the plant is driven to the desired equilibrium or is forced to follow any desired trajectory. If discontinuous control should not be used the strong connections of the energy shapping plus damping injection method with the method of the equivalent control are demonstrated and the choice of continuous controllers is also fully justified. 3) The naturality of the approach comes from differential flatness. We introduce flatness from a good number of practical examples where the system is shown to be differentially flat. The passivity + sliding mode controller design is now complemented via a differential flatness computed nominal trajectory. Dynamical parameterizations, which degenerate into static parametrizations for equilibrium conditions, are naturally provided by the flatness property of the given system. The parametrization allows for assessing the possibilities of indirect control of non minimum phase outputs. This dynamical parametrization also allows for the unequivocal determination of all basic elements needed in the passivity based controller design technique. The minimum phase or nonminimum phase character of particular outputs, as well as the zero state detectability properties, hence possibilities of static output feedback strategies, are all determined from the differential parametrization provided by the flatness of the system. Hence, the trajectory planning stage of the controller design is approached in terms of the most natural output exhibited by the system: the flat outputs, which are related to every variable in the system and describe the intrinsec linear controllability of the given plant. We show that the problem of nonminimum phase outputs largeley dissapears when flatness is invoked for the stabilization of the plant or the planning of the stabilizing motions. At this stage we, specifically, do not insist on error trajectory linearization. Flatness, which is a strong structural property, is exploited at the design state without the need of destroying useful nonlinearities. However, linearizing strategies are also explored due to their inherent simplicity and ease of design using standard pole placement.
Professor Carlos de Souza Robust State Estimation for Uncertain Linear Systems One of the fundamental problems in control systems is the estimation of the state variables of a dynamic system using available noisy measurements. Estimation methods in the minimum variance (or H-2) sense, i.e. the celebrated Kalman filtering, and in the minimax sense, i.e. H-infinity filtering, have been developed in the past decades. These methods rely on the knowledge of a perfect dynamic model for the signal generating system in order to provide a guaranteed performance. In many cases, however, only an approximate model of the system is available and, in such situations, these methods can fail to provide an acceptable performance. This lecture is concerned with the problem of robust state estimation for linear systems subject to parameter uncertainty in the matrices of the system state-space model. Design methods of robust filters with an optimized guaranteed performance in the H-2 or H-infinity sense, in spite of large parameter uncertainty, will be discussed.
Professor Dawn Tilbury Design and Analysis of Networked Controllers for Mechanical Systems As computers and computer networks continue to decline in price, distributed control systems are becoming more common. Each module of a mechanical system can have its own control system, with local I/O, computing power, and control algorithm. The mechanical modules interact physically while the control modules communicate through a network. This distributed architecture allows subsystems to be designed and tested independently. A distributed architecture is also more fault-tolerant---if one controller fails, the rest of the system may continue working in a degraded fashion. The subsystems may also be re-used in other systems, such as in reconfigurable manufacturing systems. However, most standard control designs are based on a centralized strategy, with one control algorithm supervising the entire system, and point-to-point connections between the sensors and actuators. There are several challenges that must be addressed when building a networked control system. It is well-known that the network communication induces an unavoidable time delay in the control system, which can degrade the performance or even destabilize the system. We will present a formal characterization of the time delays that can be expected in a control system using standard networks. We will also present a methodology for designing a networked control system implementation architecture given the expected communication and computation delays. As the scale of a distributed system increases, the communication load on the network may result in intolerable time delays for the control system. We will discuss some results on reducing the amount of communication required by using local estimators, and analyze the tradeoffs inherent in a networked control system between sample time and network bandwidth. The talk will conclude with a discussion of ongoing and future work in this area. Reconfigurable Logic Control for High Volume Manufacturing Systems Automatic manufacturing systems with dedicated and integrated material handling can produce large quantities of high quality parts rapidly. A discrete event supervisory control system, called a logic controller, coordinates the parallel and synchronized operation of the various machines in the manufacturing system. In current industrial practice, logic controllers are programmed in a low-level language by experienced control engineers. Although each program is fairly simple at a low level, the complexity can be enormous-it is not uncommon to find systems with 10,000 or more I/O points (events). Half of total time and cost of a new manufacturing system may be attributed to the control system; this cost can be justified if the same product will be produced for ten or more years. However, as product lifecycles decrease and product varieties increase, new methods for rapidly configuring and reconfiguring high volume manufacturing systems must be developed. This talk will overview the logic control problem for high volume manufacturing systems, and present some possible solutions using formal methods from discrete event systems. Issues associated with industrial implementation will be discussed, and examples will be drawn from the automotive and shoe manufacturing industries.
|
 |