Identifying the temporal variation of nonlinear seismic site response by Artificial Neural Network

Abstract

3135 strong motion records at 50 KiK-net stations in northeastern Japan are collected. The soil site response to strong motions is classified as linear, transition between linearity and nonlinearity, and nonlinear, by the dynamic soil parameters PGA_rotD50 and I_γ. 326 records in the nonlinear set are divided into the training set (200 records) and the test set (126 records) for an artificial neural network (ANN). This ANN is constructed to identify the temporal variation of nonlinear seismic site response automatically, by combining with the soil dynamic parameters PGA_rotD50 and I_γ in moving time window. The accuracy of soil nonlinearity identified using this network reaches 98%, which is higher than the result of visual identification, and there is no overfitting.