Seismic signal denoising using variational mode decomposition and a denoising convolutional neural network

Abstract

Effectively recovering signals buried in noise remains a challenging topic in seismic data denoising. Many conventional methods often fail to accurately capture the characteristics of seismic signals. To address this issue, this study proposed an effective method called variational mode decomposition (VMD)–denoising convolutional neural network (DnCNN). The method first applies VMD to decompose the originally noisy signal into multiple intrinsic mode functions (IMFs) with band-pass characteristics, thereby achieving effective decoupling of different frequency components and noise separation. Selected IMFs are then combined into a multi-channel input and fed into the DnCNN for end-to-end modeling and denoising reconstruction. By decomposing the noisy signal into IMFs corresponding to specific frequency bands and learning them through DnCNN, the network can better extract features within each frequency band. Serving as a front-end filter, the VMD module enhances the network’s ability to represent effective frequency components, suppresses high-frequency random noise, and improves the resolution of weak signals. Experimental results demonstrated that the proposed method effectively captures signal characteristics and recovers signals from both real and synthetic seismic data. In conclusion, the proposed VMD–DnCNN method provides a robust and efficient solution for seismic signal denoising.