An Integration of Case-Based and Model-Based Reasoning and its Application to Physical System Faults

Case-Based Reasoning (CBR) systems solve new problems by finding stored instances of problems similar to the current one, and by adapting previous solutions to fit the current problem, taking into consideration any differences between the current and previous situations. CBR has been proposed as a more robust and plausible model of expert reasoning than the better-known rule-based systems. Current CBR systems have been used in planning, engineering design, and memory organization. There has been minimal work, however, in the area of reasoning about physical systems. This type of reasoning is a difficult task, and every attempt to automate the process must overcome the problems of modeling normal behavior, diagnosing faults, and predicting future behavior. CBR systems are currently quite difficult to compare and evaluate, because there is currently no common mathematical framework in which the systems can be described. The only avenue available at present for comparison and evaluation of CBR systems requires an intellectual synthesis of the semantics of the program sources. Important constraints on the operation of a CBR system are often hidden in obscure programming tricks in the system's source code. This thesis presents a hybrid methodology for reasoning about physical systems in operation. This methodology is based on retrieval and adaptation of previously experienced problems similar to the problem at hand. In this methodology the ability of a CBR to reason about a physical system is significantly enhanced by the addition to the Case-Based Reasoner of a model of the physical system. The model describes the physical system's structural, functional, and causal behavior. Additionally, this thesis presents a mathematical formalization of the case-based reasoning paradigm and a formal specification of the interaction of the CBR component with the model-based component of a case-based system. To prove the feasibility and the merit of such methodology, a prototypical system for dealing with the faults of a physical system has been designed and implemented. Through testing has been proved that this hybrid methodology allows the generation of diagnoses and prognoses that are beyond the capabilities of current reasoning systems.