Resaeach Of Earthquake Location With A Fully Convolutional Neural Network

Earthquake early warning is a earthquake prevention and disaster reduction technology that can effectively reduce the casualties and property losses caused by the earthquake.Earthquake location is a basic problem in seismology and a key link in earthquake warning.Fast and accurate earthquake location is an important prerequisite for the success of earthquake early warning.The machine learning model constructed by the convolutional neural network mimicking the visual perception mechanism of organisms is widely used in image recognition and detection.In recent years,convolutional neural networks and full convolutional neural networks have been applied to seismic location research.Compared with the traditional seismic localization method,the method can extract the feature information directly from the original waveform,and does not depend on the accurate arrival time of P wave,and the positioning result has higher precision.In this paper,the method of full convolutional neural network is adopted,and the time-frequency data obtained by the short-time Fourier transform of the waveform is used as the input of the network to locate the seismic event.The main work done in this paper is as follows:(1)We collected 1088 earthquake events recorded in the Taiwan earthquakedetecting network in 2012-2018,obtained accurate source locations by proofreading.Considering the data quality,we selected 272 earthquake events to use for the experiment of the fully convolutional neural network.For the collected and filtered waveform data,the arriving time of the P wave phase is automatically picked up by the STA/LTA method and the AIC method,and the waveform is intercepted every 2.5s from 10 s before arriving of P to the 10 s after the arriving of P wave.The waveform is subjected to short-time Fourier transform and modulo operation to obtain time-frequency data as the input,and a three-dimensional array containing the real source information is used as a label to generate an event sample.Using cross-validation,272 seismic events were divided into 16 subsets,each containing 17 seismic events.(2)We construct a full convolutional neural network,set each layer of convolutional layer,pooling layer and deconvolution layer according to the form of input data and the problem to be solved,and adopt the mean square error loss function and introduce L2 regularization term.Then we use small Batch gradient descent method and Adam optimization algorithm to train the model,dropout and shuffle are also used.We use TensorFlow,the deep learning framework,to train the full convolutional neural network and test on test events.The union of 15 subsets was taken as the training set each time,and the remaining 1 subset was used as the test set.A total of 16 tests were performed and tested for all tests.The resulting positioning error is statistically and analyzed.(3)Some parameters in the full convolutional neural network localization method are discussed.Gaussian distribution radius,convolution kernel size,initial learning rate,regularization weights and other parameters are selected.We control the other parameters unchanged,and study the influence of each parameter change on the test results of the final training model.At the same time,according to the delay of the station and the influence of the clock difference,some stations are tested after delay processing,and the test results still have good accuracy,which proves that the full convolutional neural network positioning method is suitable for the situation of a certain range of delay and clock difference.