Rapid earthquake loss updating of spatially distributed systems via sampling-based bayesian inference

Abstract

Within moments following an earthquake event, observations collected from the affected area can be used to define a picture of expected losses and to provide emergency services with accurate information. A Bayesian Network framework could be used to update the prior loss estimates based on ground-motion prediction equations and fragility curves, considering various field observations (i.e., evidence). While very appealing in theory, Bayesian Networks pose many challenges when applied to real-world infrastructure systems, especially in terms of scalability. The present study explores the applicability of approximate Bayesian inference, based on Monte-Carlo Markov-Chain sampling algorithms, to a real-world network of roads and built areas where expected loss metrics pertain to the accessibility between damaged areas and hospitals in the region. Observations are gathered either from free-field stations (for updating the ground-motion field) or from structure-mounted stations (for the updating of the damage states of infrastructure components). It is found that the proposed Bayesian approach is able to process a system comprising hundreds of components with reasonable accuracy, time and computation cost. Emergency managers may readily use the updated loss distributions to make informed decisions.