Diagnosing faults using knowledge about malfunctioning behavior

Abstract: Second generation expert systems presume support for deep reasoning, i.e. the modelling of causal relationships rather than heuristics only. Such an approach benefits from more extensive inference power, improved reusability of knowledge and a better potential for explanations. This thesis presents a method for diagnosis of faults in technical devices, which is based on the representation of knowledge about the structure of the device an the behavior of its components. Characteristic for our method is that components are modelled in terms of states related to incoming and outgoing signals, where both normal and abnormal states are described. A bidirectional simulation method is used to derive possible faults, single as well as multiple, which are compatible with observed symptoms.The work started from experiences with a shallow expert system for diagnosis of separator systems, with a main objective to find a representation of knowledge which promoted reusability of component descriptions. The thesis describes our modelling framework and the method for fault diagnosis.Our results so far indicate that reusability and maintainability is improved, for instance since all knowledge is allocated to components rather than to the structure of the device. Further more, our approach seems to allow more reliable fault diagnosis than other deep models, due to the explicit modelling of abnormal states. Another advantage is that constraints do not have to be stated separately, but are implicitly represented in simulation rules.