Graphical Inference Methods for Fault Diagnosis based on Information from Unreliable Sensors

In this paper, we study the application of decoding algorithms to the multiple fault diagnosis (MFD) problem. Prompted by the resemblance between graphical representations for MFD problems and parity check codes, we develop a suboptimal iterative belief propagation algorithm (BPA) that is based on the graphical inference method for low density parity check codes. Our simulation results suggest that the algorithm performance strongly depends on the connection density and the reliability of the alarm network. In particular, when the connection density is low and when the alarms and/or connections are unreliable, the algorithm performs almost optimally, i.e., it converges to the solution with the highest posterior probability most of the times. We also provide analytical bounds on the performance of the algorithm for special classes of systems in our framework