Convolutional variational autoencoder for ground motion classification and generation toward efficient seismic fragility assessment

This study develops an end-to-end deep learning framework to learn and analyze ground motions (GMs) through their latent features, achieve reliable GM classification, selection, generation of simulated motions. The is composed analysis workflow that transforms reconstructs GMs short-time Fourier transform (STFT), encodes decodes features convolutional variational autoencoder (CVAE), classifies generates by grouping interpolating variables. A benchmark established confirm the minor difference between original corresponding reconstructed accelerograms. encoded space reveals certain variables are directly linked dominant physical GMs. Resultantly, clustering using k-means algorithm successfully into different groups vary in earthquake magnitude, soil type, field distance, fault mechanism. By linearly two parent variables, generated with consistent class information matching response spectra. Furthermore, seismic fragility models developed for a steel frame building concrete bridge sets Using five classified, top-ranked motions, regardless recorded or accelerograms, can reasonable efficient estimates compared case adopts 230 proposed addresses compelling questions regarding assessment: How many sufficient what types should be selected.