Resilient SOP-Based ML Classification of Harmful Fiber Vibrations Buried in Realistic Broadband Acoustic Backgrounds

Paper i proceeding, 2026

Modern optical networks underpin global telecommunications and are increasingly vulnerable to sabotage attempts. Beyond data transmission, deployed fibers can act as distributed sensors, with State of Polarization (SOP) analysis emerging as a key sensing modality for detecting malicious activity on operational infrastructure. A critical open challenge is classifying harmful vibration types, such as the 80 Hz excavator harmonic that precedes fiber cuts, when they are spectrally buried in non-stationary broadband acoustic backgrounds at varying intensity levels. We address this scenario on a jacketed patch cable using a 10-classifier Machine Learning (ML) benchmark per threat class, achieving 99.40% accuracy on the single-tone (80vb) class and 99.73% on the dual-tone (80vb+135vb) class with Histogram Gradient Boosting (HGB), confirming background-invariant threat-type discrimination across the full Signal-to-Noise Ratio (SNR) range.