Neural-network-based Seismic Phase Automatic Pickup Method

Earthquake warning technology is an effective means for reducing earthquake disasters developed in the past 20 years.When earthquakes occur,in order to carry out earthquake relief work at the first time,it is necessary to dynamically monitor the aftershock sequence information collected by each seismic station in real time.The basic principle of earthquake warning technology is: using the characteristics of P wave faster than S wave,the first incoming P wave is quickly picked up and analyzed,so as to achieve early warning of the more destructive S wave and reduce the number of casualties,loss of property.Therefore,how to quickly and accurately pick up the Pwave S-wave phase is essential.The traditional methods of seismic phase picking are manual identification and STA/LTA methods.The manual identification method is time-consuming and laborious.The STA/LTA method is mainly based on signal processing methods.However,the factors affecting seismic identification are very complex and there is also noise interference.As a result,the identification error is relatively large.Therefore,it is of great significance to research and develop a new earthquake phase picking method.This paper attempts to combine the neural network method to the seismic phase identification.The advantages of using the neural network method to solve the seismic phase identification problem are as follows:(1)The seismic wave record is a random event waveform,and the neural network is a highly nonlinear network.The method has a highly non-linear mapping function,and has obvious advantages in the separation of arbitrary complex surfaces.Therefore,using a neural network can well learn the laws in random seismic waveforms.(2)In the experiments we using neural networks need not artificially understand the complex relationships among feature attributes.Without calculation formulas and constraints,they can converge to an optimal solution,and can adapt to experimental data in different earthquake environments.(3)The efficiency and accuracy of the neural network are able to solve the problem of low efficiency and low accuracy in the conventional seismic phase picking method.The main research contents and progress of the thesis include:1.Study the characteristics of seismic pick-up related features.Analyze the correlation between each attribute of the seismic signal and the time of arrival of the Pwave and S-wave.Finally,we select the time,amplitude,polarization angle,degree of polarization,ratio of vertical energy to total energy,and STA/LTA as attributes,and enter them into the neural network.2.Analyze and study the time-dependent characteristics of seismic signals.After analyzing the advantages,disadvantages and application scenarios of various neural network models,this paper choose LSTM neural network for the seismic phase picking experiments.Through the construction of a special neural network structure,it realizes the modeling and analysis of seismic time series signals,and the comprehensive analysis of temporally adjacent phases,which greatly improves the accuracy of picking up P-wave S-wave phase.3.Study and design the neural network model structure of earthquake pick-up.By using the methods of weight adjustment and data grouping,we solve the unbalanced samples problem in the experiment.Through repeated training experiments,the network structure and parameters were adjusted,and the result is finally convergent.4.With the development of verification system,we verifies the experimental results and completes the comparison experiment.The verification system is developed under the Tensor Flow neural network framework,completes the neural network phase identification training and testing experiments.Also the comparison experiments of the traditional seismic phase identification methods are completed and we analysis the experimental results.