IDENTIFICATION OF SEISMIC SPECTRAL RATIOS BY THE DEEP LEARNING

Abstract

 We propose a method for efficiently identifying observation points through ground motion’s spectral ratios using deep learning with convolutional neural networks (CNN). In this method, the observed acceleration spectrum is converted into a color spectrum associated with the amplitude. First, random waves were applied to multiple one-dimensional numerical model grounds, and the applicability of this method was examined using the acceleration Fourier spectrum of the ground surface response converted into a color spectrum. As a result, the model grounds were identified with an accuracy of 99% or more. Next, when using the H/V spectral ratios of seismic ground motions of less than 50 gal obtained at eight K-NET stations, the stations could be identified with an accuracy of 95% or more. The misclassified seismic motions had some characteristics such as long hypocentral distances. Furthermore, when the microtremor H/V spectrum ratio was input to the trained CNN, the accuracy was as low as about 50% on average with large dispersion.