基于对比学习CUTmodel网络的地震随机噪声压制

Random Noise Suppression Method of Seismic Data Based on Contrastive Learning-CUT Model Network

本文提出一种基于对比学习CUTmodel网络的地震随机噪声压制方法,CUTmodel网络架构基于最大化输入输出之间的相似信息进行对比学习,大大缩短了网络训练时间。网络损失函数由生成对抗损失和对比学习损失两部分组成,生成对抗损失保证生成数据与无噪数据更相似,对比学习损失保证生成数据尽可能保留有效信号,同时也防止生成器进行不必要的更改,提高了网络训练的稳定性和准确性。通过简单模型和Marmousi模型的CUTmodel去噪试验及其与CycleGAN和常规去噪方法的对比,验证了本文方法的有效性。最后,本文对实际数据进行去噪,同样获得较高的信噪比。

Random noise often exists in measured seismic data,which can reduce the imaging accuracy of seismic data.Therefore,random noise suppression is an important link in determining the accuracy of seismic exploration.This article proposes a seismic random noise suppression method based on contrastive learning CUTmodel network.The CUTmodel network architecture is based on maximizing the similarity information between input and output for contrastive learning,greatly reducing network training time.The network loss function consists of two parts:generation of confrontation loss and contrastive learning loss.Generation of confrontation loss ensures that the generated data is more similar to the noiseless data,and contrastive learning loss ensures that the generated data retains effective signals as much as possible,while preventing unnecessary changes to the generator,which improves the stability and accuracy of network training.The effectiveness of the proposed method was verified through CUTmodel denoising experiments using simple model and Marmousi model,as well as comparisons with CycleGAN and conventional denoising methods.Finally,denoising the actual data also achieves a high signal-to-noise ratio.