长短期记忆网络在P波初至震相识别中的实验研究  

Experimental Study on Long Short-term Memory Networks for Identifying P-wave Primary Phase

作　　者：王天哲 张万佶 祁善博 江国明 WANG Tianzhe;ZHANG Wanji;QI Shanbo;JIANG Guoming(Key Laboratory of Intraplate Volcanoes and Earthquakes(China University of Geosciences,Beijing),Ministry of Education,Beijing 100083,China;School of Geophysics and Information Technology,China University of Geosciences(Beijing),Beijing 10083,China)

机构地区：[1]陆内火山与地震教育部重点实验室(中国地质大学,北京),北京100083 [2]中国地质大学(北京)地球物理与信息技术学院,北京100083

出　　处：《CT理论与应用研究(中英文)》2025年第2期205-215,共11页Computerized Tomography Theory and Applications

基　　金：地球深部探测与矿产资源勘查国家科技重大专项(2024ZD1000100)。

摘　　要：初至震相的识别是地震数据处理中的基本内容。由于人工识别效率较低,且受到人为主观因素的影响,因此近年来陆续发展出许多自动识别初至震相的方法。然而,这些自动识别方法主要基于背景噪声和地震信号的差异,并且通常需要一个阈值,因此难以在复杂的地震区域实施或应对海量的地震数据。为克服这些不足,本文搭建7层基于长短期记忆网络(Lstm)的卷积循环神经网络,开展P波初至震相识别的实验研究,并利用南加州公开的数据集对新建的卷积循环神经网络进行训练和测试。通过与传统的卷积神经网络、自动识别算法、Pick-Net、EQtransformer网络等进行对比,本研究搭建的卷积循环神经网络的识别精度相对较高,因此可直接使用地震波形数据作为时间序列进行训练。此外,虽然本研究建立的卷积循环神经网络只有7层网络,但基本达到复杂网络模型的震相识别精度,充分说明卷积循环神经网络的优势。综上,本研究提出的基于时间序列卷积循环神经网络为P波初至震相的自动识别提供一种新思路,为快速精准的自动识别震相问题提供技术支持。Identifying primary phases of seismic waveforms is a routine task in seismic data processing.Owing to the low efficiency of manual identification and the influence of human subjective factors,many methods for the automatic identification of the primary phase have been developed in recent years.Most of these methods determine the arrival time based on the ratio between ambient noise and seismic signals.However,they typically require a threshold value,making their implementation in complex seismic regions and handling massive seismic data challenging.In this study,a seven-layer convolutional recurrent neural network based on long shortterm memory(LSTM) network was constructed,and an experimental study was conducted to identify the P-wave primary phase.The network was trained and tested using a data set from Southern California.Compared with the traditional convolutional neural network,automatic identification algorithm,Pick-Net,and EQtransformer network,the recognition accuracy of our new convolutional recurrent neural network is relatively higher;therefore,the seismic waveform data can be directly used as a time series for training.Additionally,while the new convolutional recurrent neural network has only seven network layers,it achieves an accurate phase identification of complex network models,showcasing the strengths of convolutional neural networks.In summary,our study presents a convolutional recurrent neural network based on the LSTM,offers a new idea for the automatic identification of the primary phase,and provides technical support for the rapid and accurate automatic identification of the seismic phase.

关 键 词：深度学习 初至震相 卷积循环神经网络 长短期记忆网络 时间序列 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] P315[自动化与计算机技术—控制科学与工程]