基于CycleGAN的地震数据去噪方法  

Seismic Data Denoising Method Based on CycleGAN

作　　者：傅鹏 宋晓霞 FU Peng;SONG Xiaoxia(School of Coal Engineering,Shanxi Datong University,Datong 037000,China;School of Computer and Network Engineering,Shanxi Datong University,Datong 037000,China)

机构地区：[1]山西大同大学煤炭工程学院,山西大同037000 [2]山西大同大学计算机与网络工程学院,山西大同037000

出　　处：《电子科技》2025年第4期25-30,65,共7页Electronic Science and Technology

基　　金：山西省自然科学基金(201901D111311);山西省高质量发展研究课题(SXGZL202302)。

摘　　要：针对实际地震数据被大量随机噪声干扰而难以获得配对的无噪数据问题,文中提出一种基于CycleGAN(Cycle Generative Adversarial Network)的地震数据随机噪声压制方法来获得高质量的地震数据。将残差网络引入循环生成对抗网络的生成网络中,通过跳跃连接形式加快网络的训练速度,并扩充残差块中的卷积层,增强残差块结构来更好地获取样本特征。对合成数据和实际数据分别进行实验,利用SNR(Signal to Noise Ratio)和MSE(Mean Square Error)等评价指标验证其去噪效果,并将结果与CNN(Convolutional Neural Network)去噪方法进行对比。结果表明,相较于CNN,所提方法的SNR、MSE和PSNR(Peak Signal-to-Noise Ratio)在合成数据实验中分别提升了0.59 dB、23.72、2.81 dB,在实际数据实验中分别提升了4.63 dB、1.13、0.77 dB,训练时间缩短约58%。In view of the problem that the actual seismic data is interfered by a large amount of random noise and it is difficult to obtain paired noise-free data,this study proposes a random noise suppression method of seismic data based on CycleGAN(Cycle Generative Adversarial Network)to obtain high-quality seismic data.The residual network is introduced into the generative network of cyclic generative adversarial network,and the training speed of the network is accelerated by jumping connection,and the convolution layer in the residual block is expanded and the structure of the residual block is enhanced to obtain the sample features better.Experiments are conducted with synthetic data and actual data respectively,and evaluation indexes such as SNR(Signal to Noise Ratio)and MSE(Mean Square Error)are used to verify the denoising effect.The results show that compared with CNN,the SNR,MSE and PSNR(Peak Signal-to-Noise Ratio)of the proposed method increased by 0.59 dB,23.72 and 2.81 dB respectively,in the synthetic data experiment.In the actual data experiment,the increase is 4.63 dB,1.13 and 0.77 dB,respectively,and the training time is reduced by about 58%.

关 键 词：地震数据 随机噪声 去噪 生成对抗网络 CycleGAN 图像处理 卷积神经网络 深度学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]