Automatic Earthquake Detection Method And Its Application On The Basis Of Convolutional Neural Network

Automatic detection of earthquake and precise picking of phase arrivals are key steps in seismic data processing,and are the basic data for accurate earthquake location and research on the internal structure of the earth.In the past few decades,with the increase of seismic stations and earthquake frequency,manual analysis of seismic events has been unable to meet the needs of rapid processing of seismic data.Because the frequency of microseismic events is high and the signal-to-noise is relatively low,it is easy to be mistaken for noise and miss detection.In addition,in the study of earthquake tomography,the accuracy of seismic phase arrival is more stringent.In order to better detect small earthquakes,it is urgent to develop automatic processing methods.In recent years,the use of deep learning algorithms to detect earthquakes has developed rapidly.However,there are few studies on data processing flow and neural network parameter adjustment.In this paper,Taking 8321 local earthquake data observed by Xichang array as an example,this thesis introduces in detail the data processing flow of earthquake detection with deep convolution neural network,such as data preprocessing,model training,waveform length,network layers,learning rate and probability threshold on the detection results.Then we detect the continuous waveform with the optimal model.Our research shows that data preprocessing,data augmentation can improve the detection accuracy and anti-interference ability of the model.The length of waveform window used for model training can be approximated to the maximum value from arrival time difference between S-and P-wave.The detection results of different network layers?5?8 layers?are similar.For seismic detection,it is more appropriate to set the learning rate as 10^-4?10^-3.The earthquakes detected by convolution neural network is related to the probability threshold.By drawing the tradeoff curve of precision with recall rate,it can provide a reference for selecting the appropriate probability threshold.In addition,comparative analysis using data from the Capital Circle shows that the length of the window affects the accuracy of the detection results.Applying the model trained in the Capital Circle data and the model trained in the Wenchuan aftershock data to the Xichang test set shows that the accuracy rate is reduced,indicating that the generalization of the model needs to be improved.The detection results of Phase Net,RNN and AR-Picker are compared with a reference of manually picked arrivals.Results show that Phase Net and RNN outperformed AR-Picker in terms of arrival-time picking accuracy,suggesting that the deep learning algorithm is a promising method to accurately pick arrival-time.Phase Net has the highest accuracy in arrival-time picking.Specifically,the mean difference of the P-phase arrival-time by manually picking and Phase Net is 0.06 s,and the standard deviation is 0.11 s,while the mean difference of the S-phase is 0.13 s,and the standard deviation is 0.16 s.The accuracy of arrival-time picking by RNN is slightly worse,which may be related to the small number of the training set used in this study.However,this could be greatly improved by increasing the number of the training data set and selecting appropriate parameters.