Denoising method for Φ-OTDR systems based on deep non-negative matrix factorization and non-local means filtering

• Published in: Optics communications Vol. 596; p. 132420
• Main Authors: Zhang, Shihe, Cheng, Yafeng, Ming, Changpeng, Wang, Chenxu, Wang, Hanyong, Qian, Lei, Dong, Lei, Luo, Ming, Liu, Wu, Li, Hanbing, Huang, Tianye, Li, Xiang
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2025
• Subjects: Deep non-negative matrix factorization; Distributed optical fiber sensing; Non-local means filtering; Rayleigh backscattering; Signal denoising; Φ-OTDR; Non-local means filtering; Rayleigh backscattering; Signal denoising; Φ-OTDR; Distributed optical fiber sensing; Deep non-negative matrix factorization
• ISSN: 0030-4018
• DOI: 10.1016/j.optcom.2025.132420

The phase-sensitive optical time-domain reflectometry (Φ-OTDR) system based on Rayleigh backscattering (RBS) features high spatial resolution, long sensing distance, and strong capability for continuous monitoring, offering significant application prospects in the field of distributed optical fiber sensing. In practical applications, this system is often affected by various types of noise, primarily including laser phase noise, detector thermal noise, and environmental interference, all of which seriously impact the detection and localization accuracy of weak signals. To address these issues, this study proposes a novel denoising method that combines Deep Autoencoder-like Nonnegative Matrix Factorization (DANMF) with Non-local Means (NLM) filtering. The DANMF algorithm first decomposes the RBS signal into multiple hierarchical feature representations through multilayer nonnegative transformations, providing an initial modeling of complex Rayleigh scattering signals. Then, each extracted channel feature is individually processed using NLM filtering, which further suppresses residual noise while preserving key signal details. Experimental validation on a typical Φ-OTDR device demonstrates that the proposed DANMF-NLM method significantly improves the signal-to-noise ratio (SNR) and outperforms conventional methods. Moreover, compared to traditional deep learning models, this method requires fewer labeled samples and less computational resources, making it more practical and applicable for real-world engineering scenarios with complex noise environments.