一种基于条件对抗网络的自监督地震随机噪声压制方法  

Self-supervised seismic random noise attenuation based on conditional adversarial networks

作　　者：石战战 黄果[1] 庞溯[2,3] 王元君 周强 池跃龙[2,3] SHI ZhanZhan;HUANG Guo;PANG Su;WANG YuanJun;ZHOU Qiang;CHI YueLong(School of Artificial Intelligence,Leshan Normal University,Leshan 614000,China;The Engineering and Technical College,Chengdu University of Technology,Leshan 614000,China;College of Geophysics,Chengdu University of Technology,Chengdu 610059,China;School of Land and Resources,China West Normal University,Nanchong 637002,China)

机构地区：[1]乐山师范学院人工智能学院,乐山614000 [2]成都理工大学工程技术学院,乐山614000 [3]成都理工大学地球物理学院,成都610059 [4]西华师范大学国土资源学院,南充637002

出　　处：《地球物理学进展》2023年第1期242-253,共12页Progress in Geophysics

基　　金：国家科技重大专项课题(2016ZX05026-001);四川省教育厅项目(16ZB0410);川西南空间效应探测与应用四川省高等学校重点实验室开放基金(YBXM202102001)联合资助。

摘　　要：针对地震数据标注困难,提出基于改进的条件对抗网络的自监督随机噪声压制方法.训练过程分为2步:(1)向合成地震记录混入随机噪声构造含噪声-纯净训练集,采用监督学习策略,通过改进的条件生成对抗网络学习地震数据的有效特征;(2)借助自监督损失函数,利用目标域实际数据对预训练模型进行微调.2步训练法利用了源域合成地震记录与目标域实际地震数据之间的相似性,将源域学习到的模型迁移到目标域,实现地震数据自适应盲去噪.理论模型和实际地震数据试算结果验证所提方法具有较好的应用效果.Aiming at the difficulty of labeling seismic data,an improved self-supervised random noise suppression method based on conditional adversarial network is proposed.The training process is divided into two steps:(1)Random noise is mixed into the synthetic seismograms to construct paired noise-pure training sets.Supervised learning strategy is adopted to learn good representations of the seismic data through improved conditional adversarial network.(2)With the help of the self-supervised loss function,the practical data of the target domain is used to fine-tune the pre-training model.The 2-step training method takes advantage of the similarity between the synthetic seismograms in the source domain and the practical seismic data in the target domain,transfers the model learned from the source domain to the target domain,and realizes the adaptive blind denoising of seismic data.The experimental results of the synthetic seismograms and practical seismic data verify that the proposed method has good performance.

关 键 词：条件对抗网络 残差学习 自监督 迁移学习 随机噪声 

分 类 号：P631[天文地球—地质矿产勘探]