A Bayesian approach to online system health monitoring

This paper introduces a new online system health monitoring methodology utilizing Bayesian Belief Networks. The developed methodology enables inference with limited number of monitoring points optimally placed to obtain information on functional states of components, subsystems, and relevant physical parameters affecting the reliability of elements of the system. The approach integrates physics of failure modes when available with traditional reliability data (e.g., failures and demands) and is (1) capable of assessing current state of a system's health and probabilistic assessment of the remaining life of the system (prognosis), and (2) through appropriate data processing and interpretation can point to elements of the system that have caused or are likely to result in system failure or degradation (diagnosis). Continuous health assessment is made possible through the application of dynamic BBNs. The proposed methodology is designed to answer important questions such as how to infer the health of a system based on limited number of monitoring points at certain subsystems (“upward” inference); how to infer the health of a subsystem or component based on knowledge of the health of the main system (“downward” inference); and how to infer the health of a subsystem based on knowledge of the health of other subsystems (“distributed” inference). The methodology and algorithms are demonstrated through an example.