Structural probabilistic seismic risk analysis and damage prediction based on artificial neural network

Abstract

A novel probabilistic seismic risk analysis (PSRA) methodology based on artificial neural network (ANN) is introduced without lognormal assumption on the probabilistic seismic demand model (PSDM) and the probabilistic seismic capacity model (PSCM). The structural limit state is measured by the multidimensional limit state function. The structural damage data set is established based on incremental dynamic analysis (IDA) method, and an integrated neural network is established for damage prediction based on ten-fold cross validation. Considering the uncertainties of structural model and seismic excitation, the vulnerability curves, which reflect the upper and lower bounds of the failure probability are obtained. The mean annual frequency of exceeding a given limit state per year and the probability of exceeding a given limit state over different years are calculated based on vulnerability curve and seismic hazard function. The research shows that: The range of structural failure probability obtained by the proposed method is small, which means the established integrated neural network has strong stability and robustness. The traditional lognormal assumption may lead to inaccurate evaluation results caused by insufficient ground motion records.