Self-Supervised Denoising via Pulse Separation for Distributed Acoustic Sensing in Large-Scale Infrastructures

The fidelity of distributed acoustic sensing (DAS) data is critical for the structural health monitoring of large-scale infrastructures yet remains susceptible to complex background noise. To overcome the limitations of traditional filters at low signal-to-noise ratios (SNRs) and the 'ground truth' scarcity inherent in supervised learning, this article introduces a self-supervised denoising framework using pulse separation. Distinct from conventional approaches, our method uses the independence of noise across separated pulse intervals to formulate a self-supervised learning objective, eliminating the need for manual labeling or clean reference labels. We further incorporate a residual deep network integrated with an attention mechanism to adaptively distinguish key strain events from random noise. Laboratory experiments and field tests demonstrate that the proposed method outperforms traditional filters and state-of-the-art self-supervised models in terms of noise suppression and signal preservation, highlighting its significant potential for stability monitoring in complex engineering environments.