Gradual vs. binary conflicts in Bayesian networks applied to sensor failure detection

Bayesian Networks have various applications including medical and technical diagnosis, financial scoring, and target behavior/pattern recognition. Bayesian Classification Networks fuse evidence from heterogeneous and homogeneous sources and calculate classification results. For many reasons, pieces of evidence from different sources can carry apparently contradicting information and in these cases are called conflicting evidence. Diagnostic sensor failure-tests for application in Bayesian classification processes may be based on a binary conflict definition or a gradual conflict-level measure. This paper investigates four different failure-tests: (1) binary Conflict Binomial, (2) binary Conflict Ratio, (3) gradual Average Conflict, and (4) gradual Gauss Conflict, with (3) and (4) being new failure-test proposals. In a comparative air surveillance simulation, the detection performance of these diagnostic sensor failure-tests is evaluated and compared.