Convolutional neural network-based seismic response prediction method using spectral acceleration of earthquakes and conditional vector of structural property

Abstract

This study proposes a method for predicting the seismic response of structures using seismic information and structural properties. In the proposed method, the relationship between seismic and structural characteristics and seismic responses was investigated using a convolutional neural network (CNN) to predict the seismic response. Spectral acceleration (Sa) calculated from the seismic wave was selected as seismic information used in CNN-based seismic response prediction techniques. The study introduced seismic information and structural properties, which correspond to the parameters to express the structure's unique characteristics or nonlinear hysteretic behaviors that determine the response characteristics of the structure subjected to seismic waves. Meanwhile, Sa and structural properties were utilized to constitute the input map of CNN and predict the maximum inter-story drift ratio that corresponds to the output of CNN. As data corresponding to the period range of interest rather than a scalar value for a specific period, Sa is rearranged in matrix form to constitute the input map of CNN. Structural properties are also placed in the input map of CNN as scalar values are converted into conditional vectors. To confirm the validity of the proposed method, multiple CNN-based models with changes in the information of the input map are presented, and their prediction performances are compared. Furthermore, CNN-based models that additionally consider seismic intensity measures are presented, and their influences on seismic response prediction performance are analyzed. In addition, a vast number of linear and nonlinear structures were generated, and seismic responses extracted via seismic analysis of multiple earthquakes were used to create datasets for training the presented models. The prediction performance of the presented models trained using the datasets was compared. The validity of the simultaneous use of structural properties with Sa, the introduction of conditional vectors, the additional use of seismic intensity measures, and their contributions to improving prediction performance were also examined.