叠前随机噪声深度残差网络压制方法  

Pre-stack random noise suppression with deep residual network

作　　者：李海山[1] 陈德武[1] 吴杰[1] 常德宽 LI Haishan;CHEN Dewu;WU Jie;CHANG Dekuan(Northwest Branch,Research Institute of Petroleum Exploration&Development,PetroChina,Lanzhou,Gansu 730020,China)

机构地区：[1]中国石油勘探开发研究院西北分院,甘肃兰州730020

出　　处：《石油地球物理勘探》2020年第3期493-503,467-468,共13页Oil Geophysical Prospecting

基　　金：中国石油天然气集团公司科技项目“深层及非常规物探新方法新技术”(2019A-3312);中国石油天然气股份有限公司科技项目“智能化地震噪音压制技术研究及在塔里木沙漠区的应用”(kt2017-12-03-1)联合资助

摘　　要：深度残差网络作为一种先进的深度学习算法,近年来得到学术界和工业界的高度关注。针对叠前地震记录中随机噪声的智能高效压制问题,首先根据深度残差网络原理设计了一种深层非线性去噪网络,然后利用构建的高质量随机噪声训练集对该网络进行训练,在高维空间实现对随机噪声特征的自动学习,从而拟合出含噪声地震记录与随机噪声之间的非线性映射关系,实现随机噪声自动压制。模型数据测试和实际资料应用均证明了文中方法的有效性。尽管其去噪能力与标签数据获取方法相当,但去噪效率及适应能力明显优于标签数据获取方法,为应对TB级叠前地震数据的去噪问题提供了一种可借鉴的思路。Deep residual network,as an advanced deep learning algorithm,has received high attention from academic and industrial circles in recent years.To realize intelligent and efficient suppression of random noise in pre-stack seismic records,first,a deep nonlinear denoising network is designed based on the principle of deep residual network,and then the network is trained by the constructed high-quality random noise training sets to automatically learn the features of random noises in a high-dimensional space,so as to fit the nonlinear mapping relationship between noisy seismic records and random noises,and achieve the purpose of automatic suppression of random noises.Both model test and field application have proved the effectiveness of this method.Though the denoising capability of this method is as good as the method used for generating label data,the former has better denoising efficiency and adaptability than the latter.It provides an idea to deal with the denoising problem of TB-level pre-stack seismic data.

关 键 词：深度学习 卷积神经网络 深度残差网络 随机噪声 去噪 

分 类 号：P631.4[天文地球—地质矿产勘探]