A Novel Bayesian Network Approach for Engineering Reliability Assessment

Abstract

Bayesian network (BN) learning is difficult task due to the limitation of available data in most engineering fields. On the other hand, BN which constructed bases on experts’ judgments is hard to evaluate and often inaccurate. In this work we introduce a novel method called expert-based structural EM (Ex-SEM) for Bayesian network construction which combines limited objective data and experts’ knowledge into the BN learning process. Our experimental results show that incorporating prior knowledge from experts into Bayesian learning process could help effectively discover causal relationships in the network as well as improve accuracy of learned model. Finally, Ex-SEM was adopted to assess the reliability of a Hydraulic power plant (HPP) in Taiwan. The maintenance data were collected for more than one year; nevertheless, some data were missing. To incorporate the experts’ knowledge from the HPP maintenance sections, four experts were interviewed. The performance of Ex-SEM was compared with original SEM and K2, one popular BN learning algorithm. It is found that Ex-SEM outperformed SEM and K2 in terms of accuracy and efficiency.