Expert Fault-Diagnosis Under Human-Reporting Bias

An important class of problems is the application of expert systems to fault diagnosis where sensors are reporting symptoms and the expert system uses these in a Bayes or modified Bayes mode as evidence to help compute the posterior estimate of the source and/or nature of the fault. One of the complaints of the expert-system developer-community is that Bayes formula can rarely be applied in pure form due to lack of data from which to compute the priors and likelihood ratio elements; less defensible evidential reasoning models are becoming prevalent. For some applications, sufficiently large pools of such data do exist; however, their validity is suspect due to the lack of built-in test or built-in sensor reporting. That is, these failure data-bases depend on human operator reporting of failure events and causes. This article explores human-operator-introduced validity problems as part of an attempt to develop a strategy for compensating for failure data invalidities to the point where a Bayes approach can be possible. After elaborating on the Bayes formulation, the design of the experiments are reviewed. Results are then presented and discussed along with suggestions for further research.