Interpretable denoising of distributed acoustic sensing vertical seismic profile data using adaptive consistent prior net

Abstract

Distributed Acoustic Sensing (DAS) is vital in seismic exploration, especially Vertical Seismic Profile (VSP) studies, due to its high sensitivity, fine velocity modeling, and resistance to harsh conditions. However, DAS VSP data often suffer from low quality due to complex noise sources. Previous studies show that convolutional neural network (CNN) denoising methods can achieve good results, but they often lack interpretable mechanisms for denoising, posing a challenge in understanding and interpreting the results. To overcome these limitations, we propose an adaptive consistency prior network (ACPNet) algorithm, based on a model-guided design. First, we introduce an adaptive consistency prior algorithm to form a model-based denoising method that guides the network design, leading to an end-to-end trainable and interpretable denoising network. We constructed a comprehensive training dataset, and the network is trained using the L1 loss function to achieve effective signal-to-noise separation. The combined synthetic data and field data denoising results demonstrate that the proposed method can effectively suppress various types of noise in DAS-VSP data and has a strong signal recovery capability. Furthermore, the denoising performance of the proposed method surpasses that of traditional methods and several deep learning techniques.