基于循环一致性生成对抗网络的地震数据随机噪声压制方法

压制随机噪声、提高信噪比是地震数据处理中的关键任务。为此，提出一种基于循环一致性生成对抗网络（CycleGAN）的地震数据随机噪声压制方法。构建的CycleGAN由两个生成器和两个判别器构成，为防止网络退化，生成器由Resnet构成，用以学习含噪数据与无噪数据之间的特征映射；为提高网络的分辨率和准确性，选用PatchGAN作为判别器；同时，在传统对抗损失的基础上，添加循环一致性损失，用以提升网络训练的稳定性。完成网络构建后，针对模型数据和实际数据调整网络参数，训练和测试网络，分析去噪前后数据的信噪比和均方根误差；并通过计算单道数据频谱，进一步分析局部去噪效果。模型数据和实际数据测试结果表明，该方法能够较好地消除地震数据中的随机噪声，且去噪效果优于小波阈值去噪方法，从而验证了所提方法的可行性。

Random noise suppression method of seismic data based on CycleGAN

Suppressing random noise and enhancing the signal-to-noise ratio (SNR) are the primary tasks in seismic data processing. As for random noise in seismic data, we proposed a suppression method based on cycle-consistent generative adversarial networks (CycleGAN). The CycleGAN was composed of two generators and two discriminators. We took Resnet as the generator to learn the chara-cteristic mapping between noise-containing data and noise-free data, thereby preventing network degradation, and PatchGAN as the discriminator to improve the resolution and accuracy of the network. Besides traditional adversarial loss, the cycle consistency loss was added to improve the stability of network training. After network construction, network parameters were adjusted according to theoretical and actual data for network training and testing, and the SNR and root mean square error of data before and after denoising are analyzed. In addition, the frequency spectrum of single-channel data was calculated to analyze the denoising results. The test results based on theoretical and practical data demonstrate that the proposed method can remove random noise in seismic data, with better denoising results than those of the wavelet threshold method, which proves the feasibility of the proposed method.