基于前馈去噪卷积神经网络的地震数据去噪方法  

Seismic data denoising method based on feedforward denoising convolution neural network

作　　者：王丹荔 周怀来[1,2,3] 王元君[1,2] 吕芬 何坪易 WANG Danli;ZHOU Huailai;WANG Yuanjun;LU Fen;HE Pingyi(Chengdu University of Technology College of Geophysical;Chengdu University of Technology State Key Laboratory of Oil and Gas Reservoir Geology and Exploitation;Chengdu University of Technology Key Laboratory of Earth Exploration and Information Technology of Ministry of Education,Chengdu 610059,China)

机构地区：[1]成都理工大学地球物理学院,成都610059 [2]成都理工大学油气藏地质及开发工程国家重点实验室,成都610059 [3]成都理工大学地球探测与信息技术教育部重点实验室,成都610059

出　　处：《物探化探计算技术》2023年第1期17-27,共11页Computing Techniques For Geophysical and Geochemical Exploration

基　　金：四川省科技厅重点研发项目(21ZDYF2939)。

摘　　要：随机噪声压制是提高地震数据信噪比的有效方法,这里利用前馈去噪卷积神经网络(Denoising Convolutional Neural Network,DnCNN)的深度学习去噪方法,对地震数据随机噪声的去除进行了研究,同时利用Mish激活函数构建M-DnCNN网络进一步提升网络模型的去噪性能。该方法基于神经网络与统计学原理,通过卷积神经网络自动提取特征,利用单个残差单元来预测噪声,即输入含噪地震数据,经过M-DnCNN学习后输出预测的噪声,将输入与输出噪声作差,得到去噪后的地震数据。在实验部分,首先利用合成地震数据对该方法的去噪效果进行验证,同时将M-DnCNN与均值滤波和f-x域预测滤波方法进行去噪对比,结果表明:M-DnCNN在去除地震数据随机噪声方面优于原始DnCNN,与其他两种去噪方法相比能够更有效地压制随机噪声;随后将M-DnCNN用于实际地震数据的去噪实验,实验证明M-DnCNN在实际地震数据的去噪处理中同样适用,其在保留并突出有效信号的同时可达到较好的去噪效果。Random noise suppression is an effective method to improve the signal-to-noise ratio of seismic data.This paper studies the removal of random noise of seismic data using the depth learning denoising method of feedforward denoising convolution neural network(DnCNN).At the same time,M-DnCNN network is constructed by using a mish activation function further to improve the denoising performance of the network model.This method is based on neural network and statistical principles,automatically extracts features through convolution neural network,and uses a single residual unit to predict noise,that is,input noisy seismic data,output predicted noise after M-DnCNN learning,and make a difference between input and output noise to obtain denoised seismic data.In the experimental part,firstly,the denoising effect of this method is verified by synthetic seismic data.At the same time,M-DnCNN is compared with mean filtering and f-x domain prediction filtering methods.The results show that M-DnCNN is better than the original DnCNN in removing random noise from seismic data and can suppress random noise more effectively than the other two denoising methods.M-DnCNN is used in the denoising experiment of actual seismic data.The experiment shows that M-DnCNN also applies to denoising processing of actual seismic data.It can achieve a good denoising effect while retaining and highlighting effective signals.

关 键 词：随机噪声 地震数据去噪 Mish激活函数 M-DnCNN 

分 类 号：P631.4[天文地球—地质矿产勘探]