Application Of Machine Learning In Picking Up The First Arrival Of Seismic Wave

Accurate and fast acquisition of seismic phase travel time is the basis of many kinds of seismological research,which is of great significance to the study of hypocenter localization ? waveform inversion ? travel time tomography ? rapid earthquake warning and so on.With the increase of the number of stations in the world,the total amount of seismic data increases explosively;at the same time,there are a lot of natural or artificial noises in the station records,so manual selection of travel time has been unable to meet the needs of some researches.Now there are many automatic methods to achieve this goal,but there are some limitations.It is necessary to study the accurate?fast and anti-noise method of automatic travel time pick-up.In this paper,based on supervised learning and unsupervised learning,two machine learning classification algorithms are used to pick up phase travel time,which is the boundary point between signal and noise.The classification algorithm uses the difference of frequency,energy and statistics between signal and noise to classify the waveform data.In order to test the accuracy of the new method,10000 simulation data with known accurate phase travel time are used for travel time pick-up test,and the simulation data is composed of the superposition of reyko wavelet and Gaussian white noise.The results of travel time pick-up are compared with the results of traditional automatic pick-up methods.The actual seismic data of30 high-quality stations of Hi-climb project is used to pick up P-wave travel time.When the noise level increases to the point where the signal is completely submerged by noise,the wave is processed by symbol function to explore its influence on the travel time pick-up accuracy of the new method.The results show that,compared with the traditional method,the two machine learning algorithms used in this paper can pick up the travel time more accurately.When the amplitude of the superimposed noise is large,and the machine learning classification algorithm can pick up the phase travel time accurately and quickly,which shows good anti-noise performance.In the case of strong noise,it is helpful to improve the accuracy of travel time pick-up of machine learning classification algorithm by using symbol function to process the waveform.