Bayesian Reliability Prognosis for Systems with Heterogeneous Information

Abstract

A Bayesian methodology for prognosis of system reliability with heterogeneous reliability information is presented. Available information may be in the form of physics-based or experiment-based mathematical models, historical reliability data, or expert opinion. Such information typically describes the failure rates of individual components of the system and does not provide information on dependencies between them. The Bayesian methodology presented in this paper addresses this concern by learning the conditional probabilities in the Bayes network as observations about the system are made. First, the component and system faults are defined and the failure event tree is established. Bayesian priors for probabilities of both individual failure events and the conditional probabilities between them are established using various types of experimental data, expert opinion, or simulation data. Both the priors and conditional probabilities are updated as new data is collected, leading to an updated prognosis of system reliability. The methodology is demonstrated on an automobile startup system. 1