East China Jiaotong Fixes Fiber Security's Blind Spots With 89.79% Accuracy
A Sagnac interferometer paired with phi-OTDR and a dual-branch CNN cuts false alarms to 5% on 10-km fiber security sensing
- Core result: A Sagnac interferometer added to a phi-OTDR system, paired with a dual-branch CNN, achieves 89.79% accuracy and only 5% false alarm rate on 10-km fiber.
- The problem solved: Polarization-induced fading creates blind spots in standard phi-OTDR systems that no downstream AI can recover from; the Sagnac channel fills those gaps at the physical layer.
- Benchmark contribution: A standardized, reproducible evaluation framework comparing SVMs, probabilistic classifiers, single-branch CNNs, and fusion CNNs under identical conditions, with NAR, FNR, and latency alongside accuracy.
- Channel grouping surprise: How you split 12 input channels between two CNN branches dramatically changes performance; the naive 50/50 split yielded only 51.25% accuracy vs. 78.75% for the best grouping.
- Open source: Full implementation and scripts available at github.com/wawa-abc/das.
- Tested on: 15,419 samples across 6 acoustic event classes (digging, knocking, walking, water, fence-shaking, background) on a real 10-km outdoor fiber deployment.
Fiber-optic security systems are quietly becoming the backbone of perimeter defense, pipeline surveillance, and civil infrastructure monitoring. The technology at their core, called Distributed Acoustic Sensing (DAS), turns ordinary buried fiber into a continuous microphone that can detect footsteps, digging, or fence-cutting over tens of kilometers. But the dominant sensing method has a stubborn hardware flaw that AI alone cannot fix. A new paper from researchers at East China Jiaotong University and Guangdong University of Technology attacks that flaw at the physical layer, then wraps the whole system in a reproducible benchmark that the field has been missing.
The paper introduces a Sagnac-assisted enhanced phi-OTDR architecture paired with a standardized evaluation framework for DAS event recognition. On a 10-km outdoor fiber testbed, their dual-branch fusion CNN reached 89.79% accuracy, 89.83% macro-F1, and a nuisance alarm rate of just 5.00%, outperforming every other method in the benchmark.
The flaw that kills field deployments
The standard approach, phase-sensitive optical time-domain reflectometry (phi-OTDR), works by firing laser pulses into a fiber and measuring the tiny phase shifts in the light scattered back from imperfections in the glass. Because the round-trip time of each pulse encodes position, you get a full spatiotemporal map of vibrations along the entire fiber. The problem is polarization-induced fading (PIF): the polarization state of light drifts randomly as it travels through ordinary fiber, and at certain positions the backscattered signal can drop so sharply that the SNR collapses entirely, creating blind spots.
Phi-OTDR systems remain vulnerable to polarization-induced fading, local signal degradation, and performance instability under strong environmental interference and harsh field conditions. Due to the random spatial distribution of the refractive index and the high coherence of the light source, the backscattered signal may suffer from interference fading. Fading points, characterized by extremely low intensity backscattering, pose challenges for acquiring accurate external disturbance information. When a real intrusion event happens to land on one of these fading positions, the system simply misses it.
A second set of eyes on the fiber
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
