Reasoning Systems for Diagnostics and Prognostics

Traditionally, automated diagnostic methods have been based on identifying a single parameter which characterizes the behavior of a piece of equipment in the presence of a specific fault or a group of faults. Management of the equipment is based on removal from service should the parameter rise above a given, predetermined level, identified on the basis of previous experience. It is very rare, however, that a single parameter manages to capture all the information relevant to a given fault condition, such that optimal fleet management can be achieved on this basis alone. Rather, it is much more common that relevant information is contained in a number of different parameters, each describing a different aspect of the situation. In these circumstances, the optimal maintenance strategy can only be achieved using some form of reasoning system, which indicates the most appropriate action based on some weighting of the available evidence. Many different methodologies for performing automated reasoning have been developed within the artificial intelligence community, a number of which have been exploited within systems developed for diagnostic and prognostic purposes. However little, if any, attempt has been made to understand how the differences between these methods affect the results produced, and indeed whether there is any substantive difference between them. In this paper, different automated reasoning systems which have been proposed for equipment diagnostics and prognostics applications are considered and evaluated from a practical viewpoint. The basic principles behind each method are described, as are the fundamental properties which derive from these. A comparison of how useful the different properties are likely to be in practical applications is provided, and the various problems and limitations associated with each are discussed. Finally, areas where difficulties remain are highlighted.