Research On Rapid Prediction Of Ground Motion Parameters For Earthquake Early Warning Based On Deep Learning

Ground motion parameters（peak ground acceleration,peak ground velocity and Cumulative Absolute Velocity,etc.）are highly correlated with earthquake damage and are the key basis for Earthquake early warning（EEW）to discern earthquake damage,fast and accurate prediction of them is directly related to the success or failure of EEW.The current EEW method of predicting ground motion parameters is mainly achieved by artificially selecting the feature parameters and then establishing a relationship between the features and the ground motion intensity mapping（e.g.,ground motion prediction equation）.Nevertheless,the selection of both feature parameters and mapping relationships can rely on humans empirically,which can lead to a one-sided piece and thus limit the accuracy of prediction.In comparison,deep learning can automatically learn features directly from initial arrival waves to establish mapping relationships,which can more fully utilize the information related to ground motion parameters in initial arrival waves.Thus,this paper proposes an“end-to-end”deep learning model based on the Convolutional Neural Networks（CNN）in deep learning,with the initial arrival wave of a few seconds as the input and the ground motion parameters as the output.The main work of this paper is as follows:（1）Establishing datasets,including training,validation,and testing datasets were constructed using ground acceleration records from the Ki K-net station network in Japan,and generalization datasets were constructed using ground acceleration records from the SIBER-RISK database in Chile.The train and validation datasets are used to determine the architecture and hyperparameters of the constructed deep learning model,the test dataset is used to test the prediction of the ground motion parameters,and the generalization dataset is used to test the generalization ability of the deep learning model.（2）To improve the prediction of PGA and Sa(with periods of 0.3 s,1 s,and 3 s in EEW,an“end-to-end”deep learning model（DLA）based on multilayer CNN was constructed to predict PGA and Sa using initial arrival waves as input.The results of the test dataset show that the DLA can consistently and accurately predict PGA and Sa for EEW and that the DLA has good generalization capability,and the model trained on the Japanese dataset can be directly applied to the Chilean dataset without regional limitations.（3）To improve the prediction of PGV in EEW,a deep learning model（DLV）for predicting PGV is proposed based on multilayer CNN with the initial arrival wave as input,and has better accuracy and generalization ability compared with the commonly used displacement amplitude P_d prediction PGV performed by DLV.（4）For the current EEWs inability to predict the CAV and I_A using initial arrival waves,a deep learning model（DLCI）is proposed for predicting CAV and I_A.Given that CAV and I_A are more complex than PGA and PGV（containing process information of seismic waves）,DLCI introduces the epicenter distance,focal depth,and Vs30 as auxiliary inputs to improve the accuracy of prediction.The test results show that DLCI can consistently and accurately predict CAV and I_A for the EEW system,and that DLCI has good generalization ability.（5）To improve the effectiveness of the EEW in predicting the ground motion period parameters T_m and T_avg using initial arrival waves,a deep learning model（DLT）for predicting Tm and Tavg is proposed.The input of DLT is similar to DLCI with vertical first-to-earthquake waves and epicenter distance,source depth,and Vs30 of a single station as the input,and the results of testing and generalization show that DLT can predict T_m and T_avg consistently and rapidly for the EEW system,and DLT has good generalization ability without regional limitations.