Hybrid Bayesian Networks for Reliability Assessment of Infrastructure Systems

Abstract

Bayesian networks (BNs) facilitate the establishment and communication of complex and large probabilistic models that are best characterized through local dependences and hierarchical structures. In addition, they enable Bayesian updating of the model with new observations. This has motivated the application of BNs to the reliability assessment of large infrastructure networks. In order to make use of fast inference algorithms, previous research has mostly focused on discrete BNs. The size of the infrastructure networks that can be handled in such an approach is limited due to computational issues, and continuous random variables must be discretized. As an alternative, we propose the use of Gibbs sampling for approximate inference in such BNs. Because standard Gibbs sampling is inefficient in determining small failure probabilities, which are common in reliability problems, we introduce subset simulation, an advanced sampling technique, to BN inference. We also show how the samples from subset simulation can be used to estimate component importance measures. The approach is demonstrated by application to two road networks subjected to earthquakes.