三维地震数据频域无监督随机噪声压制方法

An unsupervised random noise suppression method in frequency domain for 3D seismic data

提高数据信噪比是地震资料处理过程中的关键环节。目前基于深度学习的降噪方法已取得较好效果。但该类方法以数据局部相似性为前提,采用时空域数据分窗进行处理,运算效率往往较低。考虑到地质结构的连续性,炮间数据具有一定的相似性,利用其同频率分量的低秩特点,设计了一种三维数据频域降秩的深度学习去噪方法。首先阐明三维数据的频域低秩原理,采用奇异值分解理论指导建立自编码神经网络;考虑频域随机噪声的分布特点,采用K-L(Kullback-Leibler)散度约束损失函数,改善了去噪效果。通过对合成记录和实际资料处理,并与多通道奇异谱分析(Multichannel Singular Spectrum Analysis,MSSA)及K-SVD(K-奇异值分解)方法对比,验证了该方法在去噪效果和计算效率等方面的优势。

Improving the signal-to-noise ratio is a key step in seismic data processing.The current deep learning-based noise reduction methods have achieved better results.However,these methods are carried out in the temporal-spatial domain based on the local similarity of the seismic data and the processing efficiency is low.In view of the lateral continuity of geological structure,the shot gathers are very similar.Thus,an unsupervised rank-reduction denoise method in frequency domain is proposed based on the low-rank feature of the same frequency component of 3D data.The low-rank principle in frequency domain of 3D data is expounded and the singular value decomposition theory is used to guide the establishment of autoencoding network;Considering the characteristics of random noise distribution in frequency domain,K-L(Kullback-Leibler)divergence is used to constrain the loss function to improve the denoising effect.The experiments on synthetic and field data verified the advantages of the proposed method in denoising performance and computational efficiency compared with the multichannel singular spectrum analysis(MSSA)and K-SVD(K-Singular Value Decomposition)methods.