ABSTRACT Distributed acoustic sensing (DAS) data are characterized by a low signal‐to‐noise ratio due to the complex noise present in its challenging operational environment. To enhance the quality of the DAS data, we propose a self‐supervised diffusion model to attenuate the DAS noise. Initially, the noisy data are divided into overlapping patches. The forward process of the diffusion model transforms these extracted patches to achieve a Gaussian distribution by gradually adding noise to diffuse the noisy data. In the reverse process, we utilize a customized deep learning model to predict the input noisy data, rather than the added noise. The key aspect for achieving robust denoising performance is the architecture of the deep learning network, which must possess the capability to effectively attenuate DAS noise. The customized deep learning network leverages the largest (coarsest) scale of the continuous wavelet transform (CWT) as a guiding and conditioning parameter, which enhances the reconstruction of the DAS signal and further suppresses noise. By incorporating the finest CWT scale into the diffusion model, the inference stage is directed to generate a cleaner DAS signal with improved noise attenuation. The proposed framework is evaluated using several synthetic and field data examples, demonstrating its robustness in comparison to benchmark methods.
Saad et al. (Sun,) studied this question.