应用自编码——解码架构的ConvGRU网络空变子波提取方法  

Space⁃varying wavelet extraction method by ConvGRU network with autoencoder⁃decoder architecture

作　　者：戴永寿[1] 李泓浩 孙伟峰[1] 宋建国[2] 孙家钊 DAI Yongshou;LI Honghao;SUN Weifeng;SONG Jianguo;SUN Jiazhao(College of Oceanography and Space Informatics,China University of Petroleum(East China),Qingdao,Shandong 266580,China;School of Geosciences,China University of Petroleum(East China),Qingdao,Shandong 266580,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,山东青岛266580 [2]中国石油大学(华东)地球科学与技术学院,山东青岛266580

出　　处：《石油地球物理勘探》2023年第4期753-765,共13页Oil Geophysical Prospecting

基　　金：国家自然科学基金项目“复杂衰减和噪声干扰下的时变子波提取与反褶积方法研究”(41974144);“基于深度学习的深地叠前时空域地震子波提取方法研究”(42274519);中国石油天然气股份有限公司重大科技项目“基于大数据的陆地时空域地震子波智能提取技术”(ZD2019‑183‑003);中央高校基本科研业务费专项“基于大数据的陆地时空域地震子波智能提取技术”(20CX05003A)联合资助。

摘　　要：地震勘探的成像和反演质量取决于地震子波的提取精度。地震子波在传播过程中主频和相位会改变,即子波形态会出现变化。现有的子波估计方法未能充分考虑子波的空变特性,同时传统空变子波提取方法对测井资料等先验信息依赖性较强,且受各类假设限制。为此,提出了一种使用自编码—解码架构与卷积门控循环单元(ConvGRU)网络的空变子波提取方法。该方法使用卷积运算与门控运算同步提取不同地震道子波的主频和相位变化特征;将提取的子波变化特征经编码得到特征变量,特征变量经解码器可更高效地提取横向道间和纵向时间上的变化特征。使用有限差分正演和非平稳褶积模型建立符合实际数据分布特点的训练数据;搭建自编码—解码网络并迭代训练网络,得到空变子波提取模型;使用该模型提取地震多道空变子波。数值仿真实验验证了该方法比传统子波提取方法具有更高的精度;中国西部实际地震资料处理结果表明,该空变子波提取方法具有一定的实际应用价值。The imaging and inversion quality of seismic exploration is inseparable from the accurate extraction of seismic wavelets.Meanwhile,the dominant frequency and phase of wavelets will change during propagation,causing changes in wavelet morphology.However,existing wavelet estimation methods still lack studies on wavelet spatial variability and rely on prior information such as well logging data.Therefore,this paper proposes a space‑varying wavelet extraction method that combines the autoencoder‑decoder architecture and ConvGRU network.The method combines convolution operation and gated calculation to extract the main frequency and phase features of wavelets in different traces.Then the features are encoded to obtain the feature variables which can more efficiently extract the features of different traces and different time by the decoder.Finite difference forward modeling and non‑stationary convolution model are employed to build training data consistent with actual data distribution.The autoencoder‑decoder network model is built and the training data is adopted to train the network and obtain a model for extracting space‑varying seismic wavelets.Finally,this model is leveraged to extract multi‑trace seismic wavelets.The numerical simulation results show that the proposed method is more accurate than traditional wavelet extraction methods.The processing of actual seismic data in western China proves that the method put forward in this paper is of certain practical application significance.

关 键 词：空变子波提取 门控循环单元 自编码—解码器 卷积神经网络 反褶积 

分 类 号：P631[天文地球—地质矿产勘探]