Use of Bayesian Networks as a decision support system for the rapid loss assessment of infrastructure systems

This paper presents an approach for the rapid seismic loss assessment of infrastructure systems, where all probabilistic variables are modeled through a Bayesian Network (BN). While BN-based approaches have been introduced as promising tools for the risk assessment of systems, they suffer from computational issues (i.e., combinatorial explosion) that prevent their application to large real-world networks that require accurate and complex performance indicators. Therefore, a hybrid BN method is introduced here, where a preliminary Monte Carlo simulation is performed in order to generate a dataset of component damage configurations, which is used to build a simplified BN structure with only a few selected components. The most critical components are selected thanks to an unbiased importance measure computed from a random forest classification. While the proposed approach generates an approximate BN structure that cannot provide exact probability distributions of losses, the application of Bayesian inference in a retro-analysis context (i.e., updating of loss projections given field observations immediately after an earthquake) has a lot of potential as a decision-support system for emergency responders. This method is applied to a road network in France, where evidence such as recorded ground-motions or observed damages is used to update the state of the system. The approximate BN structure has the ability to include complex system performance indicators, such as the additional travel time accounting for traffic flows. A sensitivity analysis on the component selection method and on the number of selected components demonstrates the stability of the posterior distributions, even with very few selected components.