Real Time Fault Detection and Diagnosis in Dynamic Engineering Systems Using Constraint Analysis

[摘要] This thesis describes some new ideas and a practically orientated implementation for fault detection and diagnosis in dynamic engineering systems. The method is designed for use on-line, it is model based, and is capable of coping with modelling inaccuracies, noisy measurements from the system and unmeasurable system states. The fault detection system is robust to false alarms, and the fault diagnosis system allows for the possibility that multiple faults may occur simultaneously. A number of system analysis algorithms are presented to extract various system equations from the model of the system. This means that the user need only enter one model of the whole system, and all of the analysis and equation solving is then handled by computer. The results of this analysis are then automatically encapsulated into a fault detection and diagnosis tool. This results in the automatic generation of a specific fault analysis tool for the system entered by the user. A "hypothesis prover" is developed here for the domain of dynamic systems, which is used to test hypotheses. Some of the ideas about multiple faults as developed by de Kleer & Williams and Reiter have been used, but these have been adapted to make them applicable for real-time, recursive, imprecise, diagnosis. (Diagnoses are imprecise because, due to modelling errors and noisy measurement, it is never possible to be 100% certain about anything.) When multiple faults are considered, the number of possible combinations becomes very large, 2N - 1, where N is the number of components. The computation required to prove a particular hypothesis, although not enormous, is not trivial either, making it impractical to prove a large number of hypotheses. To overcome this a method is proposed which involves just proving a subset of the possible hypotheses, and using the information obtained from these to reason about the other hypotheses. This requires much less computational power as the reasoning process is much less intensive than the proving process. This make the diagnosis of multiple faults possible in real-time. The methods developed here are tested on a real, noisy system where approximations are made when producing the systems' model. These tests show the potential of this approach to fault diagnosis.

[发布机构] University:University of Glasgow

[关键词] Mechanical engineering