A Multiway Model for Predicting Earthquake Ground Motion

Abstract

This paper develops a novel supervised method for predicting earthquake ground motions in the wavelet domain. The training input is a set of seismological predictors related to seismic source, path and local site conditions, and the training output consists of the weights from a multiway analysis of ground motions. We treat wavelet transforms of acceleration records as images and extract essential patterns from them using tensor decomposition. The decomposition weights of these patterns are then linked to seismological variables using general regression neural network (GRNN). The resulting nonparametric model is then used to predict the wavelet image of an accelerogram for a given set of seismological variables. The predicted image can be transformed back to the time domain using inverse wavelet transform for subsequent processing to match a given design spectrum. Unlike conventional ground motion models, the proposed approach retains the time domain characteristics of ground motions. Pearson's correlation coefficient between the vectorized forms of actual and predicted wavelet images has been used as the similarity metric in assessing the prediction capability of the resulting model. Experimental results demonstrate the ability of the proposed model to predict significant patterns in the seismic energy distribution.