Identification Of Pulselike Ground Motion Based On Machine Learning Algorithm

According to the earthquake disaster statistics and the studies,at the same magnitude and site factors,compared to the ordinary ground motions,pulselike ground motions can cause buildings damage seriously.This kind of ground motions are usually caused by the forward directivity effects.It is a fundamental task in the field of seismic engineering to accurately identify this type of ground motions.And this will help to provide a reliable pulselike and non-pulselike ground motion database for researchers to study probabilistic seismic hazard analysis and so on.In this paper,from the perspective of artificial feature extraction and representation learning,the support vector machine and the deep neural network are used respectively to identify the types of the ground motion.From the Perspectives of artificial extracted features,S transform is used to reconstruct the pulse time history in the original ground velocity record.And some features are extracted based on the reconstructed pulse time history.Then the principal component analysis is used to remove the correlation of some features and reduce their redundancy.According to the results of principal component analysis,the support vector machine is used to determine the types of ground motion.The results show that the s transform can effectively reconstruct the pulse time history in the original velocity ground motion.And this method can identify the kinds of ground motion to a certain extent.Using the support vector machine to identify the types of the ground motion,the feature extraction depends on the researchers' experience and expertise about pulselike and non-pulselike ground motion.The deep neural network can automatically extract features to identify the pulselike and non-pulselike ground motion.Therefore,from the Perspectives of representation learning,the original ground velocity ground motions are saved as image.And the image pixel vector is used as the input of the neural network.The representation of ground motions are learned by the stacked autoencoder.The softmax classifier is used to establish the corresponding deep neural network identification model.Cross validation is uesd to evaluate the model.The results show that this model has good stability and can effectively learn features to identify the types of ground motion.