Phone: +52(555)483-2020 ext. 1462
Email:
albornoz@itesm.mx
One serious problem that has haunted control engineers and artificial intelligence researchers alike from the early days on is the problem of information overload. Human plant controllers are easily overwhelmed by the sheer mass of information available to them for decision making in any real-time situation, but the same is true for automated agents operating under real-time constraints.
Large-scale systems present particular difficulties both with respect to simulation and control, and special difficulties arise when the plant to be controlled undergoes structural changes.
This thesis deals with various aspects of mixed quantitative and qualitative information processing, and tackles in particular problems related to fault monitoring, detection, characterization, isolation, and identification in large-scale systems. It addresses a number of the aforementioned difficulties, and contributes to advance the state-of-the-art in their treatment.
Problems of information overload on human operators of complex large-scale systems have been studied during the past few years mainly due to major accidents of aircrafts and in chemical, electrical, and industrial plants. Quite often, pilots/operators of such plants have had to work under immense psychological stress in situations where the onslaught of masses of unstructured information mixed with a multitude of alarms that all seemed to go off simultaneously, made an assessment of the situation under real-time constraints impossible, thereby preventing them from taking appropriate action, and yet, the consequences of a misjudgment were, at least potentially, catastrophic.
Does automatic feedback control provide us with a panacea for solving the "human overload" problem? Indeed, automatic controllers are not known to be subject to psychological stress phenomena. Yet, the same generic difficulty presents itself in automatic feedback control in a different form. Any decision-making (or control, which is the same thing) activity invariably involves solving an optimization problem. The more input variables this optimization problem contains, the higher is the dimensionality of the search space, in which the optimal solution must be found. Human plant operators are kept aware of real-time constraints, reaching a decision always in time, yet possibly making a decision that may turn out to be inappropriate since the operator didn't have enough time to explore the entire search space, thereby overlooking both better solutions and undesired side effects of the one finally selected. Automatic controllers are usually unaware of the passing of time. Hence they can be made to always reach the optimal solution, however, the process may take too long for this solution to be of any practical use. On the other hand, if a controller is being equipped with a sense of time passing, it also must limit its search, and will then begin to make the same mistakes that human operators are bound to make. Such controllers will suffer from "stress syndromes," just like any human operator would. After all, "awareness of time running out" is one of the important human "stressors."
Thus, human operators and automatic controllers need to be advised by "intelligent supervision, and/or control, and/or decision-support systems." These intelligent systems are the topic of this dissertation. They are proposed to serve as tools that may help improve the decision-making process of human operators and/or automatic controllers alike. Their basic task must be to prevent the (human or automated) decision makers from committing errors and/or from misjudging the current situation, by providing them with additional quantitative and qualitative information that can be used in the decision-making process, for detecting and discriminating faults at an as early time as possible, and for dealing with developing emergencies in an informed fashion. In particular, qualitative information has proven to be very useful in large-scale systems for discerning what is really going on, for deciding the state in which the system is at any point in time, and for assessing what would be the consequences of taking or not taking a proposed emergency action.
Following the natural order of the research, this thesis is divided in three main sections: a) Introduction, b) Methodology, and c) Applications. Each section is composed of several chapters.
The introductory section is composed of Chapters 1 and 2. This is where the problem under study and its main characteristics are defined, the use of qualitative methodologies for tackling it is justified, the application domains are delineated, and a comparison of the different qualitative modeling and simulation techniques is performed including those used in the dissertation.
In the first chapter, some important concepts were introduced, such as Large-Scale System, Variable Structure System, Intelligent Supervision and Control System, and the paradox of human vs. automatic control, with the intention of establishing the problem of information overload in large-scale systems, which has been at the origin of this research effort. The true causes of this problem and its possible solutions were stated from both the human and automatic control perspectives, with emphasis on the combination of quantitative and qualitative methodologies to solve it. The contributions of this thesis to tackling this important problem were also stated in this chapter.
Chapter 2 presents a state-of-the-art survey of fault detection and troubleshooting of dynamic systems mentioning both the quantitative and qualitative approaches. This survey provides an insight into the different methodologies (from pure diagnosis to residual generation methods in the quantitative case, and from expert systems to model-based deep reasoners in the qualitative case) used for designing Fault Monitoring Systems, and a comparison of their advantages and disadvantages when applied to non-trivial dynamic systems. In this way, the adequacy of employing qualitative methodologies to fault detection and troubleshooting is substantiated. This concludes the state-of-the-art survey of the status quo ante. All subsequent sections describe new results obtained in this research effort.
The methodological section is comprised of Chapters 3 and 5, describing Fuzzy Inductive Reasoning and Reconstruction Analysis respectively, the two methodological tools used in this thesis. Fuzzy Inductive Reasoning alone suffices to deal with problems of fault detection and troubleshooting in small- and medium-sized systems, but is incapable of tackling the information overload problem indigenous to large-scale systems. Reconstruction Analysis will be needed as an additional tool when dealing with fault detection and troubleshooting in large-scale systems, precisely for addressing the information overload problem. It seemed more appealing to keep in this thesis the same order as it was followed chronologically during the research, that is, to demonstrate the utilization of Fuzzy Inductive Reasoning in qualitative modeling and simulation as well as fault detection and troubleshooting in small- and medium-sized systems prior to introducing the additional methodological tool of the Reconstruction Analysis. This is why the two methodological chapters are not consecutive.
Chapter 3 focuses on Fuzzy Inductive Reasoning (FIR), providing full details of the development and implementation of this methodological tool. To this end, the chapter starts out with the foundations of the methodology that are rooted in General Systems Theory. It is shown how FIR can be used for qualitatively modeling and simulating continuous-time systems by means of an example (a third order linear system) that will be carried through all the phases of FIR modeling. In this chapter, the basis for a combined quantitative and qualitative modeling and simulation methodology using FIR is also provided. The chapter ends with an example of an application to mixed quantitative and qualitative modeling and simulation of a dynamic system (a hydraulic motor with a four-way servo valve). The applicability of the proposed approach to mixed quantitative and qualitative modeling and simulation is demonstrated by comparing the results of the mixed simulation with those of a purely quantitative simulation of the same system.
Chapter 5 focuses on Reconstruction Analysis (RA). As in Chapter 3, the technique is explained in full, starting out from its methodological roots in General Systems Theory. An example is carried on through the chapter in order to show the capabilities of RA to perform a causal and temporal analysis on a set of behavior variables. In the second part of this chapter, the three refinement algorithms of the Optimal Structure Analysis are applied to the same example (in the crisp and fuzzy cases), and their results are compared.
The applications section is focused on the development of a qualitative Fault Monitoring System and its application to real-world continuous-time large-scale engineering processes. This section is composed of Chapters 4, 6, and 7.
In Chapter 4, entitled "Qualitative Fault Monitoring," the necessary mechanisms used for designing a Fault Monitoring Systems are presented. The processes of causal grouping of variables, the generation of hierarchies of inductive subsystems, as well as the concepts of fault detection, identification, characterization, isolation, and diagnosis, are explained in full. Three different operating modes for the Fault Monitoring System are proposed, including one that can also be used for Variable Structure Systems. The operating modes are: Back to Training Mode, Qualitative Models Library, and Forecasting All Possible Structures. In the second part of this chapter, the Qualitative Models Library operating mode is applied to a Boeing 747 aircraft model to demonstrate its capabilities. This example also demonstrates the enhanced discriminatory power and improved forecasting capability of a fuzzy inductive reasoner over a crisp inductive reasoner. In the third part of this chapter, the Forecasting All Possible Structures operating mode is applied to the problem of structure identification in Variable Structure Systems. Two examples, a fairly simple two-water-tank system and a rather involved electric circuit example are included to demonstrate the detection, discrimination, and identification of structural changes.
Chapter 6 is focused on the selection and causal grouping of variables problem. To this end, heuristic recipes used to deal with the large number of subsystems that result from the application of the Optimal Structure Analysis technique to a large-scale system are presented. The second part of this chapter presents a comparison, from the point of view of their forecasting capabilities, between the Optimal Mask Analysis used by the Fuzzy Inductive Reasoning methodology to obtain qualitative models, and the Optimal Structure Analysis used by the Reconstruction Analysis methodology to obtain subsets of related variables that can be treated as subsystems. This comparison is made by applying both techniques to the generic third-order linear system model shown in Chapter 3 and to the Boeing 747 aircraft model that had been introduced earlier in Chapter 4.
Chapter 7 builds upon what has been developed in previous chapters, i.e., the Fuzzy Inductive Reasoning methodology (Chapter 3), the development of a Fault Monitoring System (Chapter 4), the Reconstruction Analysis methodology (Chapter 5), and the heuristic recipes for the selection and causal grouping of variables (Chapter 6). In the first part of this chapter, a combined Fuzzy Inductive Reasoning/Reconstruction Analysis (FIR/RA) methodology is proposed, and its advantages for fault detection and troubleshooting in large-scale systems are explained. In the second part of the chapter, this combined FIR/RA Fault Monitoring System is applied to a sophisticated large-scale system model, namely, a boiling water nuclear reactor model.
Finally, Chapter 8 provides a summary of the obtained results, and presents a list of open problems and possible future research efforts extending the work presented in this thesis.
Interested in reading the entire thesis? (390 pages, 3,424,483 bytes, pdf)