RSA-Net: generalized weak signal recovery for DAS VSP data in complex noise environments

Distributed acoustic sensing (DAS) technology has emerged as a leading seismic acquisition system, renowned for its high precision and efficient data collection capabilities, which are crucial for exploring deeper and more complex geological structures. The conventional signal processing techniques and the existing denoising methods are insufficient for the purpose of suppressing the various types of noise that are present in complex and diverse noise environments. This ultimately hinders the effective recovery of weak signals. To overcome these challenges, we propose the Reinforced Sparse Attention Network (RSA-Net), a deep learning framework that employs a hierarchical encoder-decoder architecture with dedicated modules for the suppression of noise and the enhancement of weak signal recovery. The network incorporates the Selective Top-k Attention (STA) module, which enables the selective focus on relevant features, and the Adaptive Mixture of Experts (AME) module, which facilitates dynamic adaptation to diverse noise types. These enhancements collectively enhance the network's generalisation capabilities across varying noise conditions. Experiments were conducted using both synthetic and field DAS VSP records, complemented by visualisation experiments that demonstrated RSA-Net's capacity to generalise across a spectrum of noise types. The results of these experiments demonstrate that RSA-Net outperforms conventional and current network-based methodologies and confirms that RSA-Net is an effective method for suppressing noise and recovering weak signals in the presence of a wide range of noise types.