应用残差网络的微地震事件五分类检测方法  

Five-category detection method for microseismic events based on residual network

作　　者：潘禹行 田宵 甘兆龙 张雄 张伟 PAN Yuxing;TIAN Xiao;GAN Zhaolong;ZHANG Xiong;ZHANG Wei(Engineering Research Center for Seismic Disaster Prevention and Engineering Geological Disaster Detection of Jiangxi Province,East China University of Technology,Nanchang,Jiangxi 330013,China;Guangdong Provincial Key Laboratory of Geophysical High-Resolution Imaging Technology,Southern University of Science and Technology,Shenzhen,Guangdong 518055,China;Shanghai Sheshan National Geophysical Observatory,Shanghai 200062,China)

机构地区：[1]江西省防震减灾与工程地质灾害探测工程研究中心(东华理工大学),江西南昌330013 [2]广东省地球物理高精度成像技术重点实验室(南方科技大学),广东深圳518055 [3]上海佘山地球物理国家野外科学观测研究站,上海200062

出　　处：《石油地球物理勘探》2024年第3期392-403,共12页Oil Geophysical Prospecting

基　　金：广东省地球物理高精度成像技术重点实验室项目“基于人工智能的地面微地震事件成像方法研究”(2022B1212010002);江西省自然科学基金项目“基于人工智能的江西地区天然地震和非天然地震事件识别方法研究”(20224BAB213047)及“多台地震实时监测的泛化神经网络及其在赣北地区的应用”(20224BAB211024);江西省防震减灾与工程地质灾害探测工程研究中心开放基金项目“基于交叉双差算法的震源位置和三维速度结构联合反演方法研究”(SDGD202210);上海佘山地球物理国家野外科学观测研究站开放基金项目“基于深度学习的地震监测和预警方法在川滇地区的应用研究”(SSOP202103)联合资助。

摘　　要：常规的微地震事件检测方法通常需要人工选取阈值,在处理大量连续记录数据时效率较低,难以适应实时监测的需求。为此,提出一种基于残差网络的微地震事件五分类检测方法,将样本分为噪声、完整的微震事件、只含有P波、只含有S波以及多个微震事件五类。该方法只需将连续记录的波形数据等分,并通过时窗调整获得完整的微震记录。通过一系列数据增广方法实现小规模实际数据样本集的模型训练,模型精度高达99%。将该方法与二分类方法同时应用于微地震监测数据检测,并通过P波、S波到时拾取和震源定位评估检测效果。研究结果表明,基于残差网络的五分类检测方法检测到了更多数量的微震事件,且具有较高的运算效率,满足实时监测的需求。Conventional detection methods for microseismic events usually require manual selection of the threshold.They are inefficient when processing a large amount of continuously recorded data and fail to meet the needs of real-time monitoring.This study proposes a five-category detection method for microseismic events based on a residual network,which divides samples into five categories:noise,microseismic events,only P waves,only S waves,and multiple microseismic events.This method only needs to equally divide the continuously recorded waveform data and obtain a complete microseismic record by shifting time windows.Through a series of data augmentation methods,the model of a small set of actual data samples is trained,and the model accuracy is as high as 99%.This method and the binary classification method are used to detect mi-croseismic monitoring data at the same time,and the detection effect is evaluated through P-wave and S-wave arrival time picking and source location.The research results show that the five-category detection method based on the residual network has greatly improved the detection quantity of microseismic events,and it has high computing efficiency,which can meet the needs of real-time monitoring.

关 键 词：微地震监测 事件检测 数据增广 残差网络 深度学习 

分 类 号：P631[天文地球—地质矿产勘探]