Use TensorFlow To Implement An Automatic Phase Picking Method Based On The Nearest Neighbor Method

This paper implements a P,S phase picking method based on the nearest neighbor algorithm proposed by Rawles et al.(2015),and applied it to the P,S phase picking method of Wenchuan and surrounding areas from July 1 to July 31,2008.Pickup of S phase.Nearest Neighbor Algorithm(KNN)is a mature machine learning method,proposed by Cover Equals in 1967.The basic idea is to determine the classification of the sample by the categories of the samples that are closest to the sample to be classified in the feature space.The method in this paper calculates a score function that depends on the Euclidean distance between the data in the window and the set of positive and negative examples through a sliding time window,and picks up the result with a specified threshold.Starting from the experimental results,this paper analyzes the station data SNR,threshold selection,positive example set structure,epicentral distance,magnitude,focal depth and other factors that may affect the number of picked results and the accuracy of the time.The results show that the signal-to-noise ratio of station data has no significant effect on the accurate picking rate of seismic catalog events;the picking threshold is negatively correlated with the picking number and the time deviation;the accuracy of the picking results quickly maintains as the size of the positive example set increases At a high level,when the size of the positive case set is greater than or equal to 10,the average accuracy of P phase pickup can reach 95%,and the S phase can reach 80%.The method in this paper is particularly stable in picking up the P phase,and its accuracy does not decrease significantly as the epicenter distance increases,the magnitude decreases or the focal depth increases;the picking accuracy of the S phase increases with the increase in the epicenter distance and the decrease in magnitude.It has declined,and when the signal-to-noise ratio of the waveform data is low,the overall performance is still stable.The development of the field of seismology and the field of machine learning mutually influence and promote each other.In recent years,many research institutions have released a variety of machine learning frameworks,including open source machine learning frameworks,which provide a lot of convenience to related researchers.This article selects the representative Tensor Flow framework released by Google,briefly analyzes the framework structure,and implements the algorithm model on this framework.The experimental results show that the static calculation graph module based on Tensor Flow can greatly improve the computing performance,up to the performance of traditional computing platforms.6 times.In recent years,broadband seismic equipment has been widely deployed,a large number of waveform records and an increasing amount of data output have made seismology have entered the era of big data.Combining the "fashionable" Tensor Flow with seismic data processing can give full play to the advantages of cutting-edge machine learning frameworks such as high performance,high scalability,and easy optimization.It is a useful attempt.