基于DBBCNN的沙漠区地震资料随机噪声衰减方法  

Research on random noise attenuation method for seismic data from deserts based on DBBCNN

作　　者：钟铁 陈云[2] 董新桐 李月 杨宝俊[5] ZHONG Tie;CHEN Yun;DONG Xintong;LI Yue;YANG Baojun(Key Laboratory of Modern Power System Simulation and Control and Renewable Energy Technolo-gy,Ministry of Education,Jilin,Jilin 132012,China;Department of Communication Engineering,Northeast Electric Power University,Jilin,Jilin 132012,China;College of Instrumentation and Electrical Engineering,Jilin University,Changchun,Jilin 130026,China;College of Communication Engineering,Jilin University,Changchun,Jilin 130012,China;College of Geo-exploration Science and Technolo-gy,Jilin University,Changchun,Jilin 130026,China)

机构地区：[1]现代电力系统仿真控制与绿色电能新技术教育部重点实验室,吉林吉林13201 [2]东北电力大学通信工程系,吉林吉林132012 [3]吉林大学仪器科学与电气工程学院,吉林长春130026 [4]吉林大学通信工程学院,吉林长春130012 [5]吉林大学地球探测科学与技术学院,吉林长春130026

出　　处：《石油地球物理勘探》2022年第2期268-278,I0001,I0002,共13页Oil Geophysical Prospecting

基　　金：国家自然科学基金重点项目“沙漠地震勘探四维非均匀波特性随机噪声数理表征及其非线性降维消噪系统”(41730422);吉林省教育厅科技项目“基于稀疏、多尺度深度学习网络的地震勘探数据高效智能去噪理论研究”(JJKH20210094KJ)联合资助。

摘　　要：由于沙漠地区采集环境恶劣、地表地质条件复杂,勘探资料信噪比普遍较低;同时,沙漠区随机噪声与有效信号存在频谱混叠现象,噪声压制难度较大,给后续反演、成像和解释等工作带来了不利影响。近年来,以去噪卷积神经网络(Feed-forward Denoising Convolutional Neural Networks,DnCNN)为代表的深度学习去噪方法已应用于复杂随机噪声抑制,但传统降噪网络一般是根据单一尺度信息提取数据特征,导致针对复杂勘探记录的处理能力可能会下降。为实现沙漠地区复杂噪声的有效衰减,提出一种新型多分支去噪卷积神经网络(Diverse Branch Block Convolutional Neural Networks,DBBCNN)。与传统的DnCNN相比,DBBCNN将不同尺度、不同复杂度的分支结合在一起,丰富了特征空间,并且采用长路径操作融合全局特征和局部特征,提升了网络针对弱信号的特征表达能力。模拟和实际数据实验结果表明,DBBCNN可有效压制沙漠地震资料中的复杂随机噪声,且处理后的记录信噪比显著提升。In the desert area of the Tarim Basin,the collected exploration data generally has a low signalto-noise ratio(SNR)due to the harsh collection environment and complex surface geological conditions.In addition,spectrum aliasing exists between random noise and effective signals,and noise suppression is challenging,which have negative impacts on subsequent procedures such as inversion,imaging,and interpretations.In recent years,deep learning denoising methods,represented by feed-forward denoising convolutional neural networks(DnCNNs),have been employed to suppress complex random noise.However,traditional denoising networks generally extract data features on the basis of single-scale information,which results in the degeneration of denoising capability when confronting complex exploration data.To achieve effective attenuation of complex noise from deserts,this paper proposed a new denoising network,namely the diverse branch block convolutional neural network(DBBCNN).Unlike traditional networks,DBBCNN combines the branches in different scales and complexity to enrich the feature space.Then,the long-path operation fuses global and local features to improve the feature expression ability of the network for weak signals.Both simulations and field experimental results show that the proposed method can effectively suppress the complex random noise from deserts with a significant increment of SNR.

关 键 词：沙漠地震 卷积神经网络 DnCNN 去噪网络 信噪比 

分 类 号：P631[天文地球—地质矿产勘探]