基于门控循环单元主导的深度网络半监督地震衰减补偿方法  

Semi-supervised seismic attenuation compensation method based on GRU dominated deep neural network

作　　者：戚倩玉 高静怀[1,2] 陈红灵 高照奇 黄研 陈娟 曹永亮 王建花 QI QianYu;GAO JingHuai;CHEN HongLing;GAO ZhaoQi;HUANG Yan;CHEN Juan;CAO YongLiang;WANG JianHua(School of Information and Communications Engineering,Faculty of Electronic and Information Engineering,Xi′an Jiaotong University,Xi′an 710115,China;National Engineering Research Center of Offshore Oil and Gas Exploration,Xi′an 710049,China;Exploration and Development Research Institute of CNPC Changqing Oilfield Branch,Xi′an 710003,China;CNOOC Research Institute Ltd.,Beijing 100027,China)

机构地区：[1]西安交通大学电子与信息学部信息与通信工程学院,西安710115 [2]海洋油气勘探国家工程研究中心,西安710049 [3]中国石油天然气股份有限公司长庆油田分公司勘探开发研究院,西安710003 [4]中海油研究总院有限责任公司,北京100027

出　　处：《地球物理学报》2023年第7期2997-3010,共14页Chinese Journal of Geophysics

基　　金：国家重点研发计划重点项目(2020YFA0713403,2020YFA0713400);中国石油长庆油田分公司“揭榜挂帅”项目“黄土塬三维高分辨率关键处理技术研究”(GS2021-01)联合资助。

摘　　要：地震波在地下传播时存在能量衰减,呈现非平稳性,导致地震资料的分辨率降低,不利于储层刻画.为了解决这个问题,需要对地震记录进行衰减补偿来消除非平稳性.考虑到GRU(Gated Recurrent Unit)即门控循环单元在处理时间序列时具有长时和短时记忆的优势,本文构建了以GRU为主导的网络结构(简称GRU网络),提出了基于GRU网络的地震衰减补偿方法.实际地震数据处理中,测井资料往往非常有限,导致标签数据较少.为解决小样本问题,本文借鉴自编码器的思想,除井旁道和测井合成地震记录的标签外,将其余地震记录作为无标签数据引入训练,半监督地学习非平稳数据到平稳数据的非线性映射,实现了地震记录的衰减补偿.最后利用含噪合成地震记录和实际地震资料测试了本文提出的方法,证明了其有效性.Seismic wave propagation in the subsurface is characterized by energy attenuation and nonstationarity,which reduces the seismic resolution and is not conducive to the fine description of reservoir characteristics.To solve this problem,it is necessary to compensate for the seismic attenuation to eliminate nonstationarity.Considering the advantages of GRU(Gated Recurrent Unit)in long-term and short-term memory when processing time sequence data,this paper constructs a neural network structure composed mainly of GUR(abbreviated as GRU neural network)and proposes a deep learning seismic attenuation compensation method based on GRU neural network.Log data is often very limited in the field seismic data processing,resulting in less labeled data.To solve the small sample size problem,inspired by autoencoder,we use synthetic seismic records and log data as labeled data and the remaining seismic records as unlabeled data for training to learn the nonlinear mapping from nonstationary data to stationary data in a semi-supervised manner to compensate for the seismic attenuation.Finally,the proposed method is tested by using noisy synthetic seismic data and field seismic data and its effectiveness is proved.

关 键 词：地震衰减补偿 深度学习 半监督学习 

分 类 号：P631[天文地球—地质矿产勘探]