Seismic Events And Phase Automatic Identification Based On Deep Learning

Introducing computer into seismic monitoring system,seismic event detection and seismic phase recognition have always been the problem that the seismologists want to overcome.With the rapid increase of seismic stations in China in recent years,seismic data has become a huge amount of data,and relying solely on manual analysis of seismic phases has been unable to keep up with data output.Due to the workload problem,manual pick-up has to discard many common seismic phases,even all the phases of a station.Therefore,automatic seismic event detection and automatic phase identification have become more and more important international hot issues.Artificial intelligence has developed rapidly in recent years.And seismologists naturally want to use artificial intelligence methods to solve the problem of automatic seismic event detection and automatic phase identification.The development of artificial intelligence benefits from the introduction of deep learning methods.This paper also attempts to find a deep learning method to explore automatic seismic event detection and seismic phase identification.Generally speaking,the problem of seismic event detection and seismic phase identification is generally to find the seismic event before identifying the seismic phase.According to this idea,this paper designs a double convolution structure model,which can be divided into two stages: seismic event detection and seismic phase identification.The models of two stages are all 9-layer convolution neural network.In the first stage,seismic events are detected from the continuous waveform data.And in the second stage,seismic phases are identified from the seismic events detected in the first stage.The double convolution structure model method is based on convolution neural network.Compared with the traditional phase identification method,it has the advantages of automatic feature extraction,no artificial threshold setting,strong model learning ability,strong non-linear expression ability,strong anti-interference ability,high efficiency and high accuracy in classifying seismic events and identifying seismic phases.Compared with other deep learning methods,the dual convolution structure model method in this paper attempts to identify seismic phase from seismic event detection based on convolution neural network,so as to explore a high-precision and efficient method for identifying seismic events and phases.In order to improve the generalization performance,anti-interference ability and phase hit rate of the double convolution structure model,and further reduce the phase error,this paper improves the sample enhancement,tuning and super parameters optimization.Four sets of sample enhancement methods,including Gaussian noise,random noise splicing,random selection of noise,and random interception of seismic events,are used to expand the training set.The expanded training set is used to train the double convolution structure model,which effectively controls the model over-fitting and improves the generalization performance and anti-interference ability of the model.The length of time sliding window,learning rate,Batch Size and other hyperparameters are optimized to further improve the accuracy of phase recognition.Through experimental comparison,the length of time sliding window of the first stage model is 30 s,and that of the second stage model is 2 s.The learning rate is 0.0001.Batch Size chooses 128.Finally,the median filtering is applied to the probability vector(The phase identification model determines the vector of each time window as the probability value of the phase arrival time)output from the second stage model.The results show that the phase error is reduced by 0.2%-0.8% and the hit rate is increased by 3%-4% after the median filtering of the probability vector.There are two stages in the work of this paper.The first stage is to explore the identification of Pg and Sg phases of a single station.The data of the three-month observation of L0230,a relatively high-visible seismic station in the Nyingchi area of Tibet,is selected as the training set.Identify data for 6 stations in the Linzhi area of Tibet for one month.In addition,Alibaba Cloud and the China Earthquake Administration jointly conducted the AI aftershock competition data set to train and identify the phase,and compare the results with other methods.The second stage of work selects the seismic network and uses multiple pieces of information to further eliminate the interference and ultimately determine the earthquake event.Using the data of 13 stations in the Yanhuai Basin and its surroundings,the apital circle network of Yanhuai Basin is a region that has been highly concerned in the metropolitan area.The academician team of Chen Yuntai and the German GFZ have established a relatively dense array here and conducted long-term observations.There far from the heavy industry area The observation conditions are relatively good.In order to verify the validity of the two-convolution structure model method,the trained model was applied to the continuous waveform data of the 13 seismic stations in the Yanhuai Basin in 2015.Compared with manual cataloging,the double convolution structure model method can detect earthquake events that are manually catalogued and missed,and can identify more phases and improve seismic cataloging.Among the earthquake events identified by three stations and three stations,the missed detection rate of the double convolution structure model method was 4.5%.In the single and two identified seismic events,the model missed detection rate was 15.9%.In addition,the method of double convolution structure model is applied to the public data of aftershock capture AI contest held by Institute of Geophysics of China Seismological Bureau and Aliyun in 2017 and the station data of Linzhi in Tibet,and good results are obtained.The experimental analysis shows that the method of double convolution structure model can effectively detect earthquake events and identify seismic phases,and the error of phase identification is small,the hit rate is high,and the missed detection rate is low.In this paper,based on the deep learning technique,this paper proposes a dual convolution structure model method,which uses different sample enhancement methods to improve the anti-interference ability and generalization performance of the model.The probability vectors obtained from the phase recognition model are filtered by median filter to reduce the error of the phase recognition and improve the correct recognition rate of the phase.Finally,automatic seismic event detection and phase recognition with high precision and accuracy are realized.