一种基于条件对抗网络的自监督地震随机噪声压制方法

Self-supervised seismic random noise attenuation based on conditional adversarial networks

针对地震数据标注困难,提出基于改进的条件对抗网络的自监督随机噪声压制方法.训练过程分为2步:(1)向合成地震记录混入随机噪声构造含噪声-纯净训练集,采用监督学习策略,通过改进的条件生成对抗网络学习地震数据的有效特征;(2)借助自监督损失函数,利用目标域实际数据对预训练模型进行微调.2步训练法利用了源域合成地震记录与目标域实际地震数据之间的相似性,将源域学习到的模型迁移到目标域,实现地震数据自适应盲去噪.理论模型和实际地震数据试算结果验证所提方法具有较好的应用效果.

Aiming at the difficulty of labeling seismic data,an improved self-supervised random noise suppression method based on conditional adversarial network is proposed.The training process is divided into two steps:(1)Random noise is mixed into the synthetic seismograms to construct paired noise-pure training sets.Supervised learning strategy is adopted to learn good representations of the seismic data through improved conditional adversarial network.(2)With the help of the self-supervised loss function,the practical data of the target domain is used to fine-tune the pre-training model.The 2-step training method takes advantage of the similarity between the synthetic seismograms in the source domain and the practical seismic data in the target domain,transfers the model learned from the source domain to the target domain,and realizes the adaptive blind denoising of seismic data.The experimental results of the synthetic seismograms and practical seismic data verify that the proposed method has good performance.