基于层序统计结构和空间地质...的深度学习高分辨率处理方法_高洋.pdf

石油科学通报 2023 年 6 月 第 8 卷第 3 期：290-302基于层序统计结构和空间地质结构的深度学习高分辨率处理方法高洋1，孙郧松2，王文闯2，李国发1*1 中国石油大学(北京)油气资源与探测国家重点实验室，北京 1022492 东方地球物理勘探有限责任公司物探技术研究中心，涿州 072751*通信作者,收稿日期:2023-03-30中国石油天然气集团有限公司科学研究与技术开发项目(2021ZG03、2021DJ1206)联合资助摘要 高分辨率地震数据在地震数据处理中扮演着关键角色，特别是当地震勘探目标变得越来越复杂时，它可以提供更准确的储层识别和描绘。近年来，随着深度学习技术的快速发展，它越来越多地被引入到高分辨率地震数据处理中。基于大量标记数据，建立了低分辨率地震数据和高分辨率地震数据之间的复杂非线性关系。然而，深度学习在高分辨率数据处理中的精度与稳定性高度依赖于训练集的准确性与多样性。深度学习技术在生产中实际应用的主要挑战之一是稀疏的井数据，这经常导致训练集受限。为了解决这个问题，本文提出了一种基于深度学习的高分辨率处理方法，通过使用大量逼真的训练集，将井数据所表示的分层结构与地震数据所表示的空间地质结构相结合。建立训练集包括三个步骤：(1)使用井数据计算波阻抗序列，并利用高斯匹配函数拟合波阻抗高频部分的振幅分布，得到一个概率密度函数，最后生成一系列符合井数据统计分布的波阻抗序列。(2)在波阻抗序列的基础上，建立二维水平分布的波阻抗模型，并逐步添加折叠变形、倾角变形和断层变形，生成包含各种地质模式的二维阻抗模型。(3)使用阻抗模型计算反射系数，然后用反射系数模型分别卷积低频和高频子波，得到训练集。通过自动生成具有地下地质信息的大量训练集，训练的网络可以估计稳定而准确的高分辨率结果。深度学习的框架由两个部分组成：提取输入数据特征的编码部分和通过提取的特征重建输出的解码部分。此外，残差模块被整合到框架中，使网络更有效地从训练集中提取特征进而提高网络性能，从而实现计算精度和效率之间更好的平衡。通过模型数据和实际数据的测试，本文提出的方法相比于传统深度学习方法对噪声具有更好的鲁棒性，可以产生更精确且横向连续性更好的高分辨率结果。关键词 深度学习；高分辨率处理；残差模块；薄层恢复；人工智能A deep learning method for high-resolution seismic processing based on a layered statistical structure and a spatial geological structureGAO Yang1,SUN Yunsong2,WANG Wenchuang2,LI Guofa11 State Key Laboratory of Petroleum Resources and Prospecting,China University of Petroleum-Beijing,Beijing 102249,China2 Research&Development Center of BGP,CNPC,Zhuozhou 072751,China引用格式：高洋,孙郧松,王文闯,李国发.基于层序统计结构和空间地质结构的深度学习高分辨率处理方法.石油科学通报,2023,03:290-302GAO Yang,SUN Yunsong,WANG Wenchuang,LI Guofa.A deep learning method for high-resolution seismic processing based on a layered statistical structure and a spatial geological structure.Petroleum Science Bulletin,2023,03:290-302.doi:10.3969/j.issn.2096-1693.2023.03.02120162023 中国石油大学(北京)清华大学出版社有限公司 291Abstract High-resolution seismic data processing plays a crucial role in the depiction and characterization of reservoir structures,especially when exploration targets become increasingly complex.In recent years,with the rapid development of deep learning tech-nology,it has been increasingly introduced into high-resolution seismic data processing.Based on a large amount of labeled data,complex nonlinear relationships between low-resolution seismic data and high-resolution seismic data are established.However,the accuracy and stability of the results generated by deep learning in high-resolution data processing highly depend on the accuracy and diversity of training sets.One of the main challenges of practical application of deep learning-based high-resolution reconstruction in production is the sparse well data,which often leads to limited training sets.To address this issue,this paper proposes a deep learning-based high-resolution processing method that integrates the layered structure represented by well data and the spatial geological structure represented by seismic data in the working area by using numerous and realistic training sets.The establishment of the training sets includes three steps.(1)Calculate the impedance sequence using well data,fit the amplitude distribution of the high-frequency part of the impedance using a Gaussian matching function to obtain a probability density function(PDF),and gen-erate a series of impedance sequences that conform to the statistical distribution of the well data.(2)On the basis of the impedance sequences,establish a two-dimensional horizontal impedance model,and gradually add folding deformation,dip deformation,and fault deformation to generate a two-dimensional impedance model containing various geological patterns.(3)Calculate the reflection coefficient using the impedance model,and then convolute the low-frequency and high-frequency wavelets with the reflection coefficient model to obtain the training sets.By automatically generating a large number of training sets with underground geological knowledge,the trained network can estimate stable and accurate high-resolution results.The framework of deep learning is composed of two parts:an encoding part that extracts features from the input data and a decoding part that reconstructs the output from the extracted features.In addition,residual modules are incorporated into the framework to enhance performance by enabling the network to learn more effectively from the training sets,resulting in a better balance between computational accuracy and efficiency.Synthetic data and field data tests show that the proposed method has better robustness to noise and can yield more accurate and laterally more consistent high-resolution results compared to traditional deep learning methods.Keywords deep learning;high-resolution processing;residual module;thin layer reconstruction;artificial intelligence doi:10.3969/j.issn.2096-1693.2023.03.0210 引言在地震勘探领域，地震数据的分辨率常受大地滤波作用和地下介质非完全弹性的影响，从而限制了对储层的准确圈定和表征。为了克服这一挑战，研究人员提出了多种方法来提高地震数据的分辨率1。Q理论被广泛应用于描述地下介质的固有衰减，衍生了多种地震衰减补偿方法，可用于提高地震记录的分辨率2-4。Braga等5采用二范数约束的小波域反Q滤波方法，以恢复稳定带宽的地震记录。Wang等6提出了一种模态相关的自适应稳定算子，该算子能够处理黏弹性补偿引起的数值不稳定性。反褶积方法在地震资料高分辨率处理中得到了广泛的应用，旨在消除地震数据中地震子波的低通滤波效应，从而拓宽频带，提高垂直分辨率。经典的最小二乘反褶积方法对噪声比较敏感，其高分辨率结果的频带