Modeling Context-Dependent Faults for Diagnosis

Most Model-based diagnosis frameworks rely on incremental probing and the assumption that faults occur independently to infer the most likely explanation for a symptom. For systems where additional sensors are unavailable or a repair action must be issued at once, these assumptions are often inadequate and dependent faults must be considered explicitly. We introduce explicit models of context-dependent component fault behavior and show that our compositional models are well-suited for the one-shot fault diagnosis of pseudo-static systems. We develop extensions to the well-known Conflict-Directed A* algorithm to infer the most-likely system state given a fixed set of observations and show that our approach complements earlier dependency models.