Dual-branch DAS denoising strategy based on SNR-aware feature fusion

Abstract In the realm of oil exploration, the effective suppression of various noises, the balancing of local and global effects, and the recovery of weak signals are relevant and vital tasks in the processing of downhole distributed acoustic sensing data. Existing denoising methods based on convolutional neural network (CNN) and Transformer generally struggle to meet these requirements due to their single feature extraction mechanisms. For instance, relying solely on local convolution or long-distant attention mechanism may not be sufficient for comprehensive feature extraction. In this paper, we propose a signal-to-noise ratio (SNR)-aware dual-branch denoising network (SNR-ADDNet) designed to dynamically enhance effective signal features derived from both CNN and Transformer architectures. Specifically, a local convolution mechanism is employed within the short-branch to capture local features in regions with high SNRs, while a long-distant attention mechanism is utilized within the long-branch to capture non-local features in other regions. This dual-branch architecture enables the network to adapt to specific characteristics of diverse regions, facilitating more precise feature extraction. To further guarantee the effective utilization of high-quality features for signal enhancement, we propose an SNR-aware Transformer. Unlike conventional approaches that incorporate all captured features indiscriminately, our method selectively considers only those features that meet a predefined SNR threshold within the attention mechanism. In addition, a visual analysis of signal features extracted from short- and long-branches reveals our network’s proficiency in integrating the strengths of CNN and Transformer to capture both detailed textures and global features. Comprehensive experiments conducted on synthetic and field seismic data demonstrate that the SNR-ADDNet exhibits robust performance in both signal recovery and noise suppression.