Dynamic Bayesian monitoring and detection for partially observable machines under multivariate observations

Abstract

Due to restrictions of mechanical structure and sensor installation, the actual status of modern engineered machine is unobservable and only partial observations can be indirectly collected. For this type of machine, existing research works used several measurable variables with fixed thresholds to monitor and detect the failures of the system, which leads to suboptimal results. This paper proposes a dynamic monitoring and detection approach for partially observed machines under discrete multivariate observations. The actual status of the partially observed machine system is modeled as a multivariate hidden semi-Markov process with unobservable operational states and an observable failure state under a general sojourn time structure. The process parameters are estimated using expectation maximization (EM) algorithm. Based on the dynamics of the hidden state of machine deterioration, a dynamic Bayesian control chart is formulated to monitor the posterior probability of system in warning state and adaptively switches the monitoring frequency according to risks of potential failures. As the posterior probability statistic exceeds a certain level, full inspection is initiated to detect the impending failures. The objective of the dynamic Bayesian control scheme is to achieve the maximum long-run system average availability, and the optimal control problem of monitoring and detection is formulated and solved by a computational algorithm in a semi-Markov decision process (SMDP) framework. The entire procedure is illustrated using multivariate data from case studies of mechanical generators and feed systems. Comparison with other advanced methods is also given, which demonstrates a considerably better performance of the proposed dynamic Bayesian approach.