深度学习在煤矿水力压裂微震检测中的应用  

Application of deep learning in microseismic detection of hydraulic fracturing in coal mine

作　　者：李昊[1,2] 高林生[3] 刘麟[2] 邵坤 LI Hao;GAO Linsheng;LIU Lin;SHAO Kun(College of Energy Science and Engineering,Xi’an University of Science and Technology,Xi’an 710054,China;School of Mine Safety,North China Institute of Science and Technology,Langfang 065201,China;School of Safety Engineering,North China Institute of Science and Technology,Langfang 065201,China)

机构地区：[1]西安科技大学能源学院,陕西西安710054 [2]华北科技学院矿山安全学院,河北廊坊065201 [3]华北科技学院安全工程学院,河北廊坊065201

出　　处：《西安科技大学学报》2023年第4期686-696,共11页Journal of Xi’an University of Science and Technology

基　　金：国家自然科学基金项目(52174111,51874133);中央高校基本科研业务费项目(3142017009)。

摘　　要：水力压裂在煤矿中被广泛应用,通常用微震检测压裂效果。为了准确识别微震的微弱波形,为后续定位、反演等波形处理奠定基础,采用深度学习卷积神经网络与图像相结合的方法,对比分析时域卷积神经网络模型、小波卷积神经网络模型、赤池信息量准则、长短时窗法4种方法识别煤层钻孔水力压裂的微弱微震数据的效果。结果表明:时域和小波卷积神经网络模型训练和测试准确率均达到99%以上,损失函数均在0.02以下;在对一个小时连续微震数据的检测中,时域模型、小波模型微震事件识别精确率分别达到100%,84%,召回率分别达到68%,57%,优于赤池信息量准则、长短时窗法的微震事件识别精确率66%,40%,召回率42%,25%;对比煤矿水力压裂微弱真实事件识别结果,时域和小波卷积神经网络模型优于赤池信息量准则和长短时窗法方法,时域模型优于小波模型。上述结果证明深度学习的卷积神经网络模型有较强微弱波形识别能力和泛化能力,是一种更优的煤矿水力压裂微震检测方法。Hydraulic fracturing is widely used in coal mines,and microseismic is usually used to detect the fracturing effect.In order to accurately identify weak microseismic waveform and lay a foundation for subsequent positioning,inversion and other waveform processing,the method of combining deep learning CNN(convolutional neural network)with image is adopted to analyze the effect of the time domain CNN model,the wavelet CNN model,AIC and STA/LTA in identifying weak microseismic data of hydraulic fracturing in coal seam drilling.The results show that the training and testing accuracy of the time-domain and wavelet convolution neural network models are above 99%,and the loss function is below 0.02.In the detection of one hour continuous microseismic data,the identification accuracy of time domain model and wavelet model for microseismic events reaches 100%and 84%respectively.The recall rate reaches 68%and 57%respectively,which is higher than AIC and STA/LTA in identifying microseismic events with accuracy of 66%and 40%,and the recall rate is 42%and 25%.In the recognition of weak real events in coal mines,time domain CNN and wavelet CNN model are superior to AIC and STA/LTA methods,and time domain models to wavelet models.The results indicate that the deep learning CNN model has a stronger weak waveform recognition ability and generalization ability than AIC and STA/LTA,a better microseismic detection method.

关 键 词：煤矿 水力压裂 微震事件 深度学习 

分 类 号：TD713[矿业工程—矿井通风与安全]