Nonparametric probabilistic seismic demand model and fragility analysis of subway stations using deep learning techniques

Abstract

A novel and computationally efficient method for developing a nonparametric probabilistic seismic demand model (PSDM) is proposed to conduct the fragility analysis of subway stations accurately and efficiently. The probability density evolution method (PDEM) is used to calculate the evolutionary probability density function of demand measure (DM) without resort to any assumptions of the distribution pattern of DM. To reduce the computational cost of a large amount of nonlinear time history analyses (NLTHAs) in the PDEM, the one-dimensional convolutional neural network (1D-CNN) is used as a surrogate model to predict the time history of structural seismic responses in a data-driven fashion. The proposed nonparametric PSDM is adopted to conduct the fragility analysis of a two-story and three-span subway station, and the results are compared with those from two existing parametric PSDMs, i.e., two-parameter lognormal distribution model and probabilistic neural network (PNN)-based PSDM. The results show that the PDEM-based PSDM has the best performance in describing the probability distribution of seismic responses of underground structures. Different from the fragility curves, the time-dependent fragility surface of the subway station shows how the exceedance probability of damage state changes over time. It can be used to estimate the escape time and thus the number of casualties in an earthquake, which are important indexes when conducting the resilience-based seismic evaluation.