The Study On On-site Peak Ground Motion Prediction Based On Machine Learning

The development of on-site earthquake warning has an extremely important role for earthquake prevention and mitigation,and at present stage,single parameters are commonly used to estimate the target parameters in in situ warning.However,there are some unsolvable problems,mainly in the form of high dispersion of statistical relations for predicting subsequent ground shaking based on a single seismic wave characteristic parameter,and some variability in the prediction results of different types of statistical relations.In this study,to address the above problems of on-site earthquake early warning and with the goal of more accurate prediction of on-site ground motion,a support vector machine model and a convolutional neural network model for predicting the intensity of the Chinese instrumental earthquakes were constructed based on artificial intelligence methods in the time window of 1?10s after the arrival of P-wave and with 1s as the time interval,respectively,using the strong motion observation data from the K-net station network in Japan.The paper mainly accomplished the following work:1.Screening and processing the strong motion data from the Japanese K-net station network,extracting eight initial characteristic parameters of seismic waves such as Pa,Pv,Pd,CAV,DI,Ia,AMmax,and IV2,establishing the linear statistical relationship between the P-wave triggered 3s time window characteristic and the PGA and PGV of the calculated Chinese instrumental seismic intensity,and the linear relationship between the single feature and the PGA and PGV is analyzed to establish the basis for the interpretable input of the artificial intelligence present ground motion prediction model based on the statistical relationship variance.2.Constructing support vector machine algorithm models to predict on-site ground motion.In order to better utilize the advantages of the SVM algorithm,eight P-wave phases of ground motion features are selected for input into the SVM model for training,and the SVM-PGA and SVM-PGV models with interpretable inputs are constructed for the prediction of on-site ground motion using the strong motion data of the Japanese Knet network.The results show that the SVM model significantly improves the phenomenon of "overestimation of small values and underestimation of large values" compared with the traditional Pd-PGA and Pd-PGV methods,and the prediction results are not affected by the epicenter distance.As the time window increases,the standard deviation of the error between the log predicted value and the log measured value gradually becomes smaller,and the prediction results gradually become better and converge to the measured results,showing a trend of gradual convergence after 6s of Pwave triggering.The SVM model has a large improvement in the accuracy of the peak prediction of on-site ground motion.3.Predicting on-site peak ground motion based on convolutional neural network.Based on the classical convolutional neural network method in the field of artificial intelligence deep learning,eight feature parameters are selected as the input of the convolutional neural network model,and the CNN-PGA and CNN-PGV models with interpretable inputs are constructed for the prediction of on-site ground motion using the strong motion data in the time window of 1?10s after the P-wave triggering of the Japanese K-net network.The results show that compared with the traditional Pd-PGA and Pd-PGV methods,the CNN models do not have the phenomenon of "large value underestimation" in predicting PGA and PGV at the time window of 3s after the arrival of P waves.With the passage of time,the standard deviation of the logarithmic error generated by the CNN model predicting the peak ground motion gradually becomes smaller and tends to converge after 6s.The accuracy of the CNN model in predicting the on-site peak ground motion is improved significantly.4.Offline simulation of seismic events.The offline simulations of several Japanese earthquakes and Chinese earthquakes and the real-time seismic event prediction results with time windows show that the support vector machine prediction model and convolutional neural network prediction model constructed in this paper have good accuracy for the peak ground motion prediction results of larger magnitude earthquakes,and the intensity results can be successfully predicted within a few seconds after the first P-wave trigger,which verifies the feasibility and practicality of the model.