Detection of failing sensors by conflicting evidence in Bayesian classification

Classification has various applications, e.g., technical monitoring, medical diagnosis, financial scoring, and pattern recognition. In a classification problem, a class out of finitely many has to be chosen on basis of observed feature values. Bayesian Networks and particularly Naïve Bayes Networks are an established approach to classification, where evidence from different sources is fused into a classification result. Sometimes pieces of evidence from different sources provide substantially different but reliable information, called conflicts. This paper proposes an approach to use conflicting evidence for performing diagnostic failure-tests of technical sources in Bayesian classification processes. Two approaches, called Conflict Ratio Failure-Test and Conflict Binomial Failure-Test are described. Based on an air surveillance scenario, a simulation is used to provide a proof of concept and to compare these two diagnostic failure-test approaches.