Bayesian Model Selection for Diagnostics

Gregory M. Provan

Abstract: Model-Based Diagnosis (MBD) addresses the task of isolating the most likely fault given a set of system measurements. The model used for diagnostics is critical to this isolation task, yet little work exists for specifying which type of model is best suited to MBD. We apply Bayesian model selection to identify the model that optimizes a diagnostics task, according to key fault-isolation metrics. We illustrate our approach using a tank benchmark system, demonstrating the trade-offs possible by using different models for this benchmark.