基于卷积神经网络的井中分布式光纤传感器地震数据随机噪声压制新技术  

New suppression technology for the random noise in the DAS seismic data based on convolutional neural network

作　　者：董新桐 李月[2] 刘飞[3] 冯黔堃 钟铁[4] DONG XinTong;LI Yue;LIU Fei;FENG QianKun;ZHONG Tie(College of Instrumentation and Electrical Engineering,Jilin University,Changchun 130026,China;College of Communication Engineering,Jilin University,Changchun 130012,China;Department of Computer and Communication Engineering,University of Science and Technology Beijing,Beijing 100083,China;Department of Communication Engineering,Northeast Electric Power University,Jilin 132012,China)

机构地区：[1]吉林大学仪器科学与电气工程学院,长春130026 [2]吉林大学通信工程学院,长春130012 [3]北京科技大学计算机与通信工程学院,北京100083 [4]东北电力大学通信工程系,吉林132012

出　　处：《地球物理学报》2021年第7期2554-2565,共12页Chinese Journal of Geophysics

基　　金：国家自然科学基金面上项目(41974143);博士后创新人才支持计划(BX2021111)共同资助.

摘　　要：分布式光纤传感器(distributed fiber-optical acoustic sensor,DAS)是一种快速发展的具有巨大应用前景的地震勘探检波器技术.实际DAS地震资料往往会受到大量强能量随机噪声的干扰,通常表现为低信噪比(signal-to-noise ratio,SNR).这一现象给接下来的成像、反演以及解释带来了巨大的困难,因此如何压制DAS地震资料中的随机噪声并提高其SNR成为一个有待解决的技术问题.卷积神经网络(convolutional neural network,CNN)已经被证明是一种有效的噪声压制工具.通常情况下,CNN需要一个理论纯净地震数据集来优化网络,这极大地限制了CNN在DAS地震资料处理中的应用.在本文中,我们采用正演模拟的方法来构建理论纯净DAS地震数据集,通过正演模型的参数多样化增强数据集的真实性,从而获得适合DAS地震资料随机噪声压制的CNN去噪模型.此外,在网络结构方面,我们利用泄漏线性整流单元作为CNN的激活函数增强训练后模型对微弱有效信号的恢复能力;在训练过程中,通过能量比矩阵调节噪声片和有效信号片之间的SNR,增强CNN去噪模型对于不同SNR的DAS地震数据的适应性.模拟和实际实验均表明本文提出的这种正演模型驱动的卷积神经网络(forward-model-actuation convolutional neural network,FMA-CNN)能够有效地压制DAS随机噪声同时完整地恢复有效信号.Distributed fiber-optical acoustic sensor(DAS)is a rapid-developing geophone technology and will has a wide application prospect in modern seismic exploration.However,due to the strong contamination by strong random noise,the seismic data received by DAS is always characterized by low signal-to-noise ratio(SNR).This phenomenon brings lots of difficulties to the following imaging,inversion,and interpretation.Therefore,how to suppress the random noise in DAS seismic data and enhance its SNR is a technical problem to be solved.Convolutional neural network(CNN)has been proven to be an effective tool for noise suppression.Generally,CNN need a theoretical pure seismic dataset to train itself;this greatly limits the application of CNN in the random noise suppression of DAS seismic data.In this paper,we utilize forward modeling method to construct the theoretical pure DAS seismic dataset,and the completeness of this dataset is greatly enhanced by the parameter diversification of forward models,so as to achieve the optimization of CNN denoising model.Furthermore,in the aspect of network architecture,leakage rectifier linear unit is introduced as the activation function of CNN to enhance the recovery ability of trained models for weak effective signals;in the training process,an energy ratio matrix is proposed to adjust the SNR between noise patches and effective signal patches,so as to enhance the adaptability of CNN denoising model to the DAS seismic data with different SNRs.Synthetic and real experiments demonstrate that this forward-model-actuation convolutional neural network(FMA-CNN)proposed in this paper can effectively suppress the DAS random noise and completely recover the effective signals.

关 键 词：随机噪声 低信噪比 分布式光纤传感器地震数据 正演模型驱动的卷积神经网络 

分 类 号：P631[天文地球—地质矿产勘探]