应用生成对抗网络的地震数据重建和去噪一体化方法  

An integrated method of seismic data reconstruction and denoising based on generative adversarial network

作　　者：张岩[1,2] 张一鸣[1] 董宏丽 宋利伟[4] ZHANG Yan;ZHANG Yiming;DONG Hongi;SONG Liwei(School of Computer&Information Technology,Northeast Petroleum University,Daqing,Heilongjiang 163318,China;Artificial Intelligence Energy Research Institute,Northeast Petroleum University,Daqing,Heilongjiang 163318,China;Key Laboratory of Networking and Intelligent Control of Heilongjiang Province,Daqing,Heilongjiang 163318,China;School of Physics and Electronic Engineering,Northeast Petroleum University,Daqing,Heilongjiang 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318 [2]东北石油大学人工智能能源研究院,黑龙江大庆163318 [3]黑龙江省网络与智能控制重点实验室,黑龙江大庆163318 [4]东北石油大学物理与电子工程学院,黑龙江大庆163318

出　　处：《石油地球物理勘探》2024年第4期714-723,共10页Oil Geophysical Prospecting

基　　金：东北石油大学特色科研团队项目“智慧油田信息处理创新团队”(2023TSTD-04)资助。

摘　　要：在实际采集过程中,受地形条件和人为因素的影响,地震数据不仅在空间上会出现采样不足或不规则的情况,而且会混入噪声,不利于后续地震数据的处理和解释。通常将重建与去噪分为两个阶段处理,这样往往会引入额外的误差。为此,文中提出了一种基于条件韦氏生成对抗网络(cWGAN)的地震数据重建去噪一体化方法,该方法研究的重点是在缺失道和噪声的混合干扰下,准确提取地震数据的有效特征。首先,以U-Net模型为基本网络结构来构建生成器模型,分级提取地震数据同相轴特征;在判别器模型中引入条件约束,引导生成器优化梯度方向。其次,建立重建和去噪误差描述模型,该模型设计了一体化损失函数,可以兼顾重建与去噪两方面的处理任务。最后,经过合成数据和实际数据测试,证明文中所提的网络模型恢复的地震数据信噪比更高且具有较强鲁棒性。During the actual acquisition process,due to terrain conditions and human factors,seismic data can suffer from spatial under sampling or irregular sampling,as well as being contaminated by random noise,which hinders subsequent processing and interpretation.Current seismic data processing methods typically separate re⁃construction and denoising into two stages,often introducing additional errors.The focus of the integrated re⁃construction and denoising method is to accurately extract the effective features of seismic data under mixed in⁃terference from missing traces and noise.This paper proposes an integrated method for seismic data reconstruc⁃tion and denoising based on conditional Wasserstein generative adversarial network(cWGAN).Firstly,a ge⁃nerator model is constructed with the U⁃Net model as the basic network structure,and the event features of seis⁃mic data are extracted.Conditional constraints are then introduced into the discriminator model to guide the gra⁃dient optimization direction of the generator.Secondly,an error description model for reconstruction and de⁃noising is established,and an integrated loss function is designed to address both tasks simultaneously.Finally,tests on synthetic and actual data demonstrate that the seismic data recovered by the proposed network model have a higher signal⁃to⁃noise ratio and good robustness.

关 键 词：地震数据处理 重建与去噪一体化 深度学习 生成对抗网络 一体化损失函数 

分 类 号：P631[天文地球—地质矿产勘探]