Study Of Seismic Wave First-to-arrival Pickup Based On Convolutional Neural Networks

Accurately picking up the first arrival time of the P wave and S wave of the seismic wave is of great significance for the precise positioning of the earthquake and the interpretation of the earthquake preparation mechanism.Relying solely on seismology experts for analysis will consume a lot of manpower and time,so there is an urgent need for an automated method to identify and pick up the first arrival of seismic waves.In recent years,deep learning has been widely used in all walks of life.Among them,the application research of convolutional neural network in the picking of seismic waves has received extensive attention from researchers.Compared with the traditional method,the convolutional neural network does not need to manually extract features,it can automatically perform feature extraction and obtain the optimal solution,and the convolutional neural network can solve the problem of large pick-up errors in traditional methods and poor model generalization.Therefore,This article conducts in-depth research based on this.In this paper,we use the seismic data from the 2020 Microsoft Innovation Cup China and the data from Shanxi stations as the data.Firstly,the data from the Microsoft Innovation Cup China 2020 seismic competition are pre-processed.Then the three-component seismic waveform data were input into the convolutional neural network for training and testing.The model was evaluated using three evaluation metrics,namely accuracy,recall and F1 score,and compared with the traditional STA/LTA and U-Net models.Finally,the generalizability of the model is verified using Shanxi station data.The main conclusions of this paper are as follows:(1)Seismic wave P-wave and S-wave first-to-arrival pickup using convolutional neural networks.Compared with the traditional STA/LTA,the convolutional neural network approach does not require manual setting of thresholds and manual selection of feature functions,relying only on the convolutional neural network to automatically extract waveform features,and the model has a higher tolerance for low signal-to-noise ratio data.(2)The R-Unet model is proposed by combining the advantages of Residual Unit and U-Net.The R-Unet model sums the downsampling part of the conventional U-Net with the Residual Unit,the pooling layer in the downsampling process is replaced by a convolution operation to reduce the loss of high-frequency information in the seismic waveform.The long connection and the residual connection are used to combine the low-level information features with the high-level information features to reduce the information loss during feature extraction and feature reduction.The R-Unet model is used to pick up the first arrivals of Pand S-waves of seismic waves,comparing the R-Unet model with the conventional STA/LTA and U-Net for analysis,the accuracy of R-Unet is 91.4% for P-wave and 84.5% for S-wave,and the recall rate is 87.4% for P-wave and 79.6% for S-wave,which are higher than the conventional STA/LTA and U-Net models.(3)The Unet++ model is improved and the Unet PP model is proposed.The Unet PP model fills the center of the traditional U-Net and integrates 1D U-Net with 1 to 4 layers,combining long and short connections to better extract seismic waveform features at different levels and scales in the seismic P-wave and S-wave,reducing the information loss coming from downsampling.The weights of noise,P-wave and S-wave are adjusted and higher weights are assigned to S-wave to increase the recognition accuracy of S-wave at first arrival.Comparing the Unet PP model with the conventional STA/LTA and U-Net for analysis,the P-wave accuracy of R-Unet is 97.7%,the S-wave accuracy is 96.4%,the P-wave recall is95.3%,and the S-wave recall is 90.4%,all of which are higher than the traditional STA/LTA and U-Net models.Although Unet PP increases the training time compared to R-Unet,the accuracy of the model for seismic wave P-wave and S-wave first-to-pickup increases,with 6.3%for P-wave and 11.9% for S-wave.(4)The generalizability of the model is verified using 179 seismic data from 8 stations in Shanxi.The results show that both the R-Unet model and the Unet PP model have small pickup errors.The R-Unet model has a P-wave initial arrival pickup error of 0.58 s and an S-wave initial arrival pickup error of 0.91 s.The P-wave initial arrival pickup error of the Unet PP model is 0.39 s,and the S-wave initial arrival pickup error is 0.46 s.The initial arrival time of P-wave and S-wave is less than 1s,and the model has good generalization to Shanxi station data.The experimental study in this paper can provide a new idea for the initial arrival time pickup of P and S waves in the future,so that the initial arrival time pickup of P and S waves can be more accurate,which is expected to provide technical support for the localization of earthquakes and the interpretation of earthquake inception mechanism.