A PROBABILISTIC THEORY OF MODEL-BASED DIAGNOSIS

Diagnosis of a malfunctioning physical system is the task of identifying those component parts whose failures are responsible for discrepancies between observed and correct system behavior. The result of diagnosis is to enable system repair by replacement of failed parts. The model-based approach to diagnosis has emerged as a strong alternative to both symptom-based and fault-model-based approaches. Hypothesis generation and hypothesis discrimination (action selection) are two major subtasks of model-based diagnosis. Hypothesis generation has been partially resolved by symbolic reasoning using a subjective notion of parsimony such as non-redundancy. Action selection has only been studied for special cases, e.g. probes with equal cost. Little formal work has been done on repair selection and verification. This paper presents a probabilistic theory for model-based diagnosis. An objective measure is used to rank hypotheses, viz., posterior probabilities, instead of subjective parsimony. Fault hypotheses are generated in decreasing probability order. The theory provides an estimate of the expected diagnosis cost of an action. The result of the minimal cost action is used to adjust hypothesis probabilities and to select further actions. The major contributions of this paper are the incorporation of probabilistic reasoning into model-based diagnosis and the integration of repair as part of diagnosis. The integration of diagnosis and repair makes it possible to troubleshoot failures effectively in complex systems.