Study On Key Technologies Of Intelligent Seismic Data Processing For Global Sparse Seismic Network

The Comprehensive Nuclear-Test-Ban Treaty monitors possible nuclear tests anywhere in the world through seismic,infrasound,hydroacoustics and radionuclide technologies.Seismic monitoring is realized through real-time recording of ground vibrations by seismic stations distributed regionally or globally.With a series of data processing,parameters such as the time,location,depth,nature and magnitude of the earthquake can be inferred.Since such method is fast sensitive and reliable,Seismic monitoring technology is widely used in many countries to monitor nuclear explosion and determine the location,nature and estimation of explosion yield.The traditional data processing of seismic monitoring is a detection-based processing method and obtains seismic event bulletin through signal detection,feature extraction,phase recognition,association and locationrealized by step-by-step serial data processing pipelines.With the dramatic increase in the number of seismic stations and the volume of historical data in the global earthquake monitoring network,the performance of traditional data processing methods is not satisfactory.New problems and challenges are emerging in the quality of automatic processing of the event bulletin and the application of historical data mining.In order to further improve the automatic processing performance of seismic data in the global sparse seismic network,especially the detection ability of weak events,this paper has carried out research on the key technologies of intelligent processing of seismic data for the global sparse seismic network,from the aspects of seismic signal detection,phase picking,as well as phase association.Incorporating machine learning and deep learning into seismic data processing,the study established a method to detect seismic signals in a specific area based on waveform characteristics,a seismic phase picking method based on deep neural networks,a seismic phase association algorithm based on Bayesian theory,and an intelligent processing model for earthquake monitoring data has been established.The main innovative results are as follows:(1)A method for detecting tele-seismic P-wave signals in specific areas based on waveform characteristics is proposed to achieve reliable detection of seismic signals in specific areas under the global sparse seismic network.First,select the seismic signal waveform characteristics,such as the peak value of STA/LTA,the ratio of STA/LTA,the average spectral energy of a specific frequency band,and the specific energy fraction of EMD,and then use the optimal hyperparameters obtained during the training of the Gaussian process with historical event data in this area,to construct a probability model of waveform features,and finally calculate the posterior probability of joint features to achieve reliable detection of seismic signals in a specific area.This method effectively improves the detection performance of seismic signals in a specific area.(2)A seismic phase picking method based on deep neural network is proposed to solve the problem of seismic phase picking under the global sparse seismic network.A multi-task convolutional neural network model is firstly constructed to realize seismic phase detection and arrival time estimation simultaneously,in which a weighted multi-class cross-entropy loss function is defined,and a joint loss function of classification and regression is designed.The network model is trained,verified,and tested using the large-scale Southern California data set.Through transferring learning and data enhancement methods,the model is transferred to a small-scale data set of stations in Northeast China,and a detection method for continuous waveform data is established.The actual data test shows that the method can accurately pick up the P and S phase of the local earthquake events in the continuous waveform data,realizing reliable detection of local events.(3)A Bayesian seismic association algorithm based on waveform envelope,seismic phase and event feature models is proposed to solve the problem of seismic phase association under the global sparse seismic network,in which the overall envelope characteristics of the waveform is introduced.A priori probability model is constructed with machine learning,and an event inference method based on Bayesian theory is designed to realize seismic phase association and event parameter inversion.Then,based on the application requirements of seismic big data processing,coupled with the efficient computation of GPU,a Bayesian seismic phase association algorithm platform is established.Through the seismic event data of Japan and its surrounding seas,the effectiveness of the feature model construction method and the seismic association algorithm is tested and verified.Finally,based on this platform,an evaluation method is designed to assess the results of NET-VISA,the IDC's next-generation seismic phase association algorithm,using the standard event bulletin.