Study On The Prediction Of Underground Motion Parameters Based On Deep Neural Network

At present,China is at the peak of the development of underground structures such as urban subways,and many cities are located in 8-degree of security.Therefore,the earthquake resistance of underground structures is of great significance for ensuring the safety of our country's engineering system.and underground motion parameters are the basis of seismic design of underground structures.The commonly used prediction methods have the problems of too large error or relatively complicated process.In recent years,many problems in the field of seismic engineering and seismology have benefited from deep learning.A series of studies such as earthquake prediction and surface motion parameter prediction have demonstrated the powerful learning and prediction capabilities of deep learning.However,there is no research on the prediction of underground seismic parameters based on deep learning.In this paper,new models are developed on the prediction of peak underground acceleration,peak underground velocity and underground spectral acceleration based on the deep neural network depend on the Ki K-net array to establish a underground motion database.The main work is summarized as follows:Firstly,The underground motion database are established taking advantage of the large network of vertical seismic arrays that have recorded many earthquakes over wide ranges of magnitudes and peak ground accelerations,which focused on 110 stations,2980 earthquakes,a total of 20256 sets of seismic records from November 18,1998 to January 6,2019,that in consideration of the completeness of geological information,the rationality of database components,and the signal-to-noise ratio of seismic records.The database is randomly divided into training data,validation data and testing data according to the ratio of 8:1:1 in view of to the distribution of PGA.Secondly,It is proposed to develop a ground-motion model for predicting underground ground motion parameters based on deep neural network,a multi-input deep neural network is proposed combining artificial neural network(ANN)with 1D convolution neural network(1D-CNN),with the feature extraction of the onedimensional convolutional neural network in terms of shear wave velocity sequence of the soil layer,whose input parameters include magnitude,depth of underground stations,ground motion parameters and soil shear wave velocity.By measuring eight evaluation indexes to ensure the consistency of the prediction results with the attenuation law of underground motion parameters,it is shown that the residuals of the prediction model basically obeyed normal distributed,the mean value of which is zero and the standard deviation less than 0.39.Finally,the result of models are compared with the result decided by onedimensional equivalent linearization method in frequency domain,the empirical formula,the seismic design requiremenrs of the underground structure.it is found that the prediction accuracy of the deep learning prediction model is higher than the above three methods,the deep learning prediction model is at least 20.63% higher than the statistical empirical formula,and 52.19% higher than the seismic design codes for underground structures.The prediction results of peak underground acceleration,peak underground velocity,and spectral acceleration are increased by 2.44%,11.27% and 4.52% respectively compared to one-dimensional equivalent linearization method in frequency domain.and avoid the process of iteratively adjustment of G and ? based on the peak strain and solve wave equations.