Research on a long-haul transmission-line distributed vibration signal denoising method based on dual-fiber sensing...

Long-haul distributed fiber-optic vibration monitoring for overhead transmission lines enables continuous sensing along the entire corridor. However, as the sensing distance increases, signal attenuation and the accumulation of environmental/system noise lead to a pronounced degradation in signal-to-noise ratio (SNR), thereby reducing the reliability of feature extraction and condition assessment. To address the difficulty of acquiring "clean-label" samples in field environments and the limited capability of conventional filtering in suppressing in-band noise, this paper proposes a dual-fiber-sensing-based weakly supervised denoising method for long-distance distributed vibration signals. By serially connecting two fibers within the same OPGW, two DAS observations corresponding to the same spatial locations are obtained, and their noise discrepancies are exploited to construct Noise2Noise (N2N) training pairs. Building on this, an N2N-Attention-U-Net network incorporating an attention-gating mechanism is developed to achieve end-to-end adaptive denoising of DAS spatiotemporal blocks. A 60-km field dataset collected from a 220-kV overhead transmission line in Anshan, Liaoning Province, China, is used to conduct comparative experiments against bandpass filtering, Wiener filtering, and the jDAS model, with evaluations performed from multiple perspectives including spatiotemporal maps, frequency-space distributions, time-domain waveforms at representative sensing points, and frequency-time evolution patterns. The results demonstrate that deep-learning-based approaches achieve superior overall performance; in particular, the proposed method substantially reduces the noise floor and stripe-like artifacts while better preserving weak spectral lines and time-varying spectral structures. Further validation on out-of-training-set samples shows that the proposed method maintains the best overall denoising performance with favorable generalization capability and robustness, providing high-fidelity preprocessed data to support subsequent condition recognition in long-distance transmission-line vibration monitoring.