Bayesian networks for disaggregation of structural reliability

Abstract

Recent developments in theory and solution algorithms have facilitated the application of Bayesian Probabilistic Networks (BPNs) to probabilistic engineering problems. BPNs’ ability to easily perform backward analysis and parameter updating offers advantage over existing methods. This paper attempts to propose a BPN for reliability assessment of structural systems based on their cut-set representation. A five-layer Bayesian network is proposed which includes random variables, component events, failure paths and failure modes to compute the structural probability of failure. Convenient inclusion of parameter dependency or correlation is an advantage of the proposed method. While in existing methods extensive computation or simplifying assumption are used to account for component and mode dependencies, the proposed method utilizes the inherent ability of BPNs to model the joint probability distributions using factors of independent conditional probabilities. This greatly reduces efforts to include parameter dependencies which may significantly affect the output. Application to a real-world example of 4-storey reinforced concrete structure is presented and optimizations to lower the computation costs are reviewed. Also it is described how disaggregation of structural failure and updated distribution of random variables can be obtained using backward inference without any additional recalculations.