基于去噪卷积神经网络的异常振幅压制方法  

Abnormal amplitude suppression method based on denoising convolutional neural network

作　　者：范承祥 郭宏伟[3] 苑益军[1] FAN Chengxiang;GUO Hongwei;YUAN Yijun(School of Geophysics and Information Technology,China University of Geosciences(Beijing),Beijing 100083,China;Weichai Power Co.,Ltd.Weifang,Shandong 261061,China;Exploration and Development Research Institute,PetroChina,Beijing 100083,China)

机构地区：[1]中国地质大学(北京)地球物理与信息技术学院,北京100083 [2]潍柴动力股份有限公司,山东潍坊261061 [3]中国石油勘探开发研究院,北京100083

出　　处：《石油地球物理勘探》2023年第4期780-788,共9页Oil Geophysical Prospecting

基　　金：国家自然科学基金项目“基于复矢量Radon变换的地震矢量波场分离方法研究”(41974157)资助。

摘　　要：地震数据中的异常振幅会造成地震道空间能量不均,导致叠前偏移出现画弧现象,从而严重干扰地震资料解释,因此压制异常振幅已成为地震资料处理中的一项重要工作。由于应用条件的限制,采用传统方法难以实现在彻底压制异常振幅的同时保护有效信号。为此,提出一种基于去噪卷积神经网络(DnCNN)的异常振幅压制方法。该方法首先根据地震异常振幅分布特点,通过网络改进与优化,搭建了适于异常振幅压制的DnCNN结构;然后采用人工合成和实际数据提取相结合的方法,制作了包含异常振幅和不含异常振幅的两种训练集;最后利用训练集对搭建的网络进行训练与学习,获得能够压制异常振幅的网络训练模型。模型数据和实际地震数据应用结果表明,该方法能够有效压制地震数据中的异常振幅,同时也保护了有效信号,与常用的传统方法相比,处理效果最佳。Abnormal amplitude of seismic data often leads to uneven spatial energy in seismic data,resulting in arc phenomenon during the prestack migration and interfering with seismic data interpretation.Therefore,suppressing abnormal amplitudes has become an important step in seismic data processing.Due to the limitation of application conditions,traditional methods fail to completely suppress abnormal amplitude while protecting effective signals.Therefore,a method for suppressing abnormal amplitude based on a denoising convolutional neural network(DnCNN)is proposed.Firstly,according to the distribution characteristics of seismic abnormal amplitude,this method builds a DnCNN structure suitable for suppressing abnormal amplitude through network improvement and optimization.Secondly,training sets with and without abnormal amplitude are produced by artificial synthesis and real data extraction.The network is trained with the training set,and a network training model that can suppress abnormal amplitude is obtained.Finally,tests on model data and real seismic data show that the proposed method can effectively suppress abnormal amplitude in seismic data,protect effective signals,and obtain better processing results than commonly used traditional methods.

关 键 词：深度学习 卷积神经网络 训练集 异常振幅 去噪 

分 类 号：P631[天文地球—地质矿产勘探]