Identification Of Pulse-like Ground Motions Based On Convolutional Neural Networks

Seismic data show that structures are more prone to severe damage under the action of pulse-like ground motion than ordinary ground motion.To avoid the subjective error of manual selection of pulse-like ground motion,quantitative identification of pulse-like ground motion has become a key research problem in the field of earthquake engineering and civil engineering.In view of the outstanding performance of convolutional neural networks in the field of artificial visual simulation,this paper explores a novel method for identifying pulse-like ground motion based on convolutional neural networks in order to overcome the problems of non-uniform parameters and conflicting recognition results in the field of pulse-like ground motion identification by traditional methods.This paper mainly includes the following aspects of research.(1)In order to obtain the weight coefficients in the training convolutional neural network to construct a robust network structure,this paper first selects a set of typical pulse-like ground motion data based on four previously proposed methods for quantitative identification of pulse-like ground motion.In view of the fact that the actual pulse-like ground motion data is relatively small,in order to form a complete set of training data,based on the characteristics of the selected typical pulse-like ground motion,the Kanai-Tajimi spectral model is used to artificially generate ground motion that are similar to the actual pulse-like ground motion characteristics.data.(2)In order to obtain the optimal network structure and data input format,this paper compares and analyzes the characteristics of 11 typical convolutional neural networks in pulse-like ground motion recognition.Finally Inception V3 is selected with the best performance for pulse recognition.Since the wavelet coefficient maps of pulse-like ground motions are significantly different from those of ordinary ground motion,the input data of this paper are two-dimensional color images of the wavelet coefficient maps of ground motions.This paper systematically discusses the influence of the effective time-history length of the wavelet coefficient map and the format of the two-dimensional image on the recognition results.(3)In order to verify the accuracy and stability of the method in this paper,this paper systematically compares the classification performance of the four new pulse recognition algorithms proposed in the past two years and the model in this paper on the corresponding data sets.According to the results of recurrence rate and further analysis of the conflicting samples,the model in this paper can effectively extract data features and accurately identify ground mot ion categories.