基于改进pix2pix GAN的多次波压制算法  

Multiple attenuation algorithm based on improved pix2pix GAN network

作　　者：张全 吕晓雨 雷芩 黄懿璇 彭博[1,2,3] 李艳 ZHANG Quan;LYU Xiaoyu;LEI Qin;HUANG Yixuan;PENG Bo;LI Yan(School of Computer Science,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Intelligent Oil and Gas Laboratory,Southwest Petroleum University,Chengdu,Sichuan 610500,China;State Key Laboratory of Oil and Gas Reservoir Geology and Exploitation(Southwest Petroleum University),Chengdu,Sichuan 610500,China)

机构地区：[1]西南石油大学计算机科学学院,四川成都610500 [2]西南石油大学智能油气实验室,四川成都610500 [3]油气藏地质及开发工程国家重点实验室(西南石油大学),四川成都610500

出　　处：《石油地球物理勘探》2024年第4期664-674,共11页Oil Geophysical Prospecting

基　　金：油气藏地质及开发工程国家重点实验室开放基金项目“石油钻井环境异常工况智能识别技术研究”(PLN2022-51);“基于高性能计算与卷积神经网络的地震多次波压制方法研究”(PLN2021-21);四川省南充市科技局开放基金项目“地震多次波高效压制与深度学习集成研究”(23XNSYSX0089)联合资助。

摘　　要：有效压制地震多次波是地震资料处理过程中的重要环节,尽管已有多种多次波压制方法,但是传统的多次波压制方法依赖先验地质构造信息且需要大量的计算,导致多次波压制效率较低,对于复杂地质条件下的多次波压制更具挑战。为此,提出将改进pix2pix GAN运用于多次波压制问题中,利用神经网络的特征学习能力,提高多次压制波的精度。这种改进的pix2pix GAN结合ResNet与U-Net作为网络的生成器,以避免深层网络引起的梯度消失或梯度爆炸现象。并在生成器中引入SE注意力机制,改进的生成器能够更好的感知地震波中一次波与多次波的特征,提升生成器性能。此外,使用多尺度判别器对更精细的地震图像细节特征和纹理信息做出真假判别。网络的输入为全波场数据,标签为一次波数据,使用两个简单地层模型和一个公开的Sigbee2B模型合成的数据集训练网络。实验结果表明,改进的pix2pix GAN比pix2pix GAN的多次波压制效果更好;网络训练一旦完成,即可有效提升多次波压制速度。The effective attenuation of seismic multiples plays a crucial role in the seismic data processing work⁃flow.Despite the existence of numerous multiple attenuation methods,traditional approaches heavily rely on prior geological structure information and require extensive calculations,resulting in slow attenuation speed.This poses an even greater challenge for multiple attenuation under complex geological conditions.To over⁃come the limitations of traditional methods and improve efficiency,this paper applies the pix2pix GAN network to the problem of multiple attenuation and utilizes the feature learning capability of neural networks to improve the processing speed.It proposes an enhanced multiple attenuation method for the pix2pix GAN network,which integrates ResNet and U⁃Net as the network generator to avoid gradient vanishing or exploding phenomena used by deep netwoorks,while incorporating the SE attention mechanism.The improved generator can better per⁃ceive the characteristics of both first⁃order and multiples,thereby enhancing its performance.Additionally,a multi-scale discriminator is employed to discern detailed features and texture information on finer seismic images for accurate identification of authenticity.The input data for the network consists of full wave field data labeled as primary wave data,with training conducted using a dataset synthesized from two simple formation models and a public Sigbee2B model.Experimental results demonstrate that the improved GAN network exhibits superior accu⁃racy in multiple attenuation compared to pix2pix GAN,effectively improving attenuation speed.

关 键 词：多次波消除 深度学习 注意力机制 ResNet Sigbee2B 

分 类 号：P631[天文地球—地质矿产勘探]