Candidate Ordering and Elimination in Model-Based Fault Diagnosis

A major step in model-based fault diagnosis is the generation of candidate submodules which might be responsible for the observed symptom of malfunction. After the candidates are determined, each subrnodule can then be examined in turn. It is useful to be able to choose the most likely candidate to focus on first so that the faulty parts can be located sooner. We propose here a systematic method for initial candidate ordering that takes into account the structure of the device and the discrepancy in outputs between the observed and expected values. We also give effective methods for a system to adjust its focus according to new information acquired during diagnosis. Under the single fault assumption, the average length of diagnosis (number of submodules evaluated) is O(logm), where m is the number of submod-ules. 1 Introduction Diagnostic reasoning based on structural and functional descriptions of a device, usually referred to as the "de-sign model", has been shown to be viable by many Al researchers [Cantone et a/. A thorough survey on model-based troubleshooting can be found in [Davis and Harnscher, 1988]. In this approach reasoning uses "first principles", i.e., knowledge of how the device works rather than knowledge of how it fails. The knowledge needed for such a system is well-structured and usually available when a device is designed. A model-based system avoids many of the difficulties of an empirical failure-based diagnosis system, e.g., in knowledge acquisition, diagnosis capability (ability to diagnose previously unseen faults) and system generalization. Since a model-based fault diagnosis system reasons directly on the structure and function of a device, it usually follows a simple control structure. It starts from the top level of the structural hierarchy of the device and tries to find outputs that violate their expectations. After detecting the violated outputs, the system uses structural descriptions to find a subset of components, which might be responsible for the observed symptom of malfunction, at the next lower hierarchical level. This step, known as candidate generation, is a major step in fault diagnosis. This process is then continued with each of the candidate components in turn until the faulty parts are found. It is desirable to have a diagnostic system which is able to choose the most likely candidate to focus on first so that the faulty parts can be located sooner. Previous work on diagnosis based on structure and behavior contains suggestions on initial candidate …