Updating conditional probabilities of Bayesian belief networks by merging expert knowledge and system monitoring data

Bayesian Belief Network (BBN) methods can be adopted for reliability analysis and real-time monitoring of critical infrastructure. Conditional Probability Tables (CPTs) are used to describe the dependencies between nodes in the BBN. The CPTs are usually defined by using either expert knowledge elicitation process or data-driven learning methods, and in particular, these methods define the CPTs statically, i.e. the CPTs are defined based only on the information available at the beginning of the analysis process. Therefore, the CPTs represent the correlation between the different components at one point in time, and potential correlations, evolving over time, are neglected. In this paper, a novel method to update the CPTs over time by merging expert judgement and the data about the system behaviour is proposed, which allows a process of continuous updating of CPTs. First of all, the method defines the CPTs by using the expert knowledge elicitation process. Then the CPTs are continuously updated whenever new data-driven evidence about the system becomes available. The proposed method is verified by assessing the health state of an in-field bridge, which experiences different magnitudes of damage. The BBN analysis of the health state of the bridge shows that the proposed method improves the accuracy of the BBN analysis, improving the diagnostics capabilities of the BBN up to 20%, by relying on a more robust and continuous updating of the CPTs