Turning Every Telecom Cable into a City-Wide Sensor Network

Introduction

This is the subject of “Leveraging Deployed Telecom Cables for Distributed Fiber Sensing Topologies and Applications,” published in the Journal of Optical Communications and Networking (Vol. 18, No. 4, April 2026). The paper was authored by Scott R. Kotrla, Principal Member of Technical Staff at Verizon, and Jeffrey A. Mundt, also of Verizon, together with Ming-Fang “Yvonne” Huang, Jian Fang, Shaobo Han, Jamie Lynn, Ezra Ip, and Ting Wang, all of NEC Laboratories America. The work reflects a collaboration between Verizon’s network engineering team and NEC Laboratories America’s optical communications research group, combining field access to live carrier infrastructure with sensing hardware and artificial intelligence (AI) expertise.

The Problem: Idle Fiber, Untapped Data

Telecom carriers have buried hundreds of thousands of miles of fiber optic cable to carry internet and phone traffic. That same glass, it turns out, is sensitive to vibration, temperature, and strain, meaning it can double as a distributed sensor stretching for miles. The technique, known as distributed fiber optic sensing (DFOS), has existed for years, but two practical limits have held it back.

The first is distance and compatibility. The most common method, called backscattering sensing, reads light that bounces backward through the fiber and works well for high-precision detection. Still, it cannot pass through the amplifiers and switching equipment (erbium-doped fiber amplifiers and reconfigurable optical add-drop multiplexers) that live telecom networks use to boost and route signals over long distances. That confines it to short, unamplified, often dark (unused) fibers, not the busy backbone links carriers operate.

The second limit is data overload. A single sensing fiber can generate enormous volumes of raw vibration data every second, far more than any human team could review, and building AI models to sort meaningful events from noise has typically required large amounts of manually labeled training data for every new fiber route. This slow and expensive process does not scale to a citywide network.

Combine Advanced Sensing with Self-Supervised AI

The authors address the distance problem by pairing two complementary sensing methods. Backscattering sensing, which reads reflected light, offers meter-level precision but cannot travel through amplified links. Forward transmission sensing, which reads light traveling in the same direction as ordinary data traffic, can pass through amplifiers and cover hundreds of kilometers. Still, it only narrows an event down to a rough stretch of fiber. The team’s hybrid approach uses forward transmission sensing to scan an entire route and flag the general area of a disturbance, then activates backscattering sensing only on that specific segment for precise, meter-level localization. It is like using a wide-angle camera to spot movement across a stadium, then zooming a second camera in on the exact seat.

To address the data and labeling problem, the researchers developed a technique called locally adaptive anomaly detection (LAAD), a self-supervised representation learning method. Instead of requiring engineers to label thousands of events on every new fiber route manually, LAAD first trains itself using location-indices as labels: predicting which location along the fiber a given signal patch came from. Because different spots along a cable have distinct, repeating vibration signatures, such as a manhole near a pump or a pole near a road, the model learns what “normal” looks like at each specific location without being told what any event is. Once trained this way, it needs far less labeled data to fine-tune for a specific downstream task like detecting cable strikes or classifying traffic.

Field Validated Across 500 Kilometers

The team tested all three sensing approaches on live carrier fiber across the Dallas-Fort Worth metropolitan area, spanning roughly 310 miles (500 kilometers) of monitored cable. In the backscattering trial, eleven sensing units deployed as a mesh network tracked cable damage over four years and, after full deployment with early warning alerts in 2024, cut the share of cable failures occurring inside monitored zones from 22% in 2023 to 9% in 2024.

In the forward transmission trial over an 80-kilometer field segment, the system detected and localized events as faint as footsteps on a manhole cover, with phase signals below 10 radians, a level of sensitivity that conventional forward transmission methods could not reliably resolve. The hybrid trial, run over a 249-kilometer, three-span link carrying live 400-gigabit-per-second data traffic, achieved a measurement resolution of 146 picoepsilon per square root hertz, a 32% sensitivity improvement over full link monitoring, while maintaining zero post-forward error correction bit errors on all 43 co-propagating data channels.

On the AI side, after self-training, LAAD classified 224 distinct fiber locations along a single route with 97.96% pretext task location-classification accuracy. The learned representation also supports cross-domain generalization: on two entirely new fiber routes, it achieved more than 97.66% downstream event-classification accuracy across a range of different classification models.

Real-World Applications

Telecom carriers, like Verizon, can use this approach to protect their own infrastructure, catching excavation and construction activity near buried cables before a backhoe causes an outage. City transportation departments and municipal planners could tap the same fiber for traffic density and road condition monitoring without installing new cameras or roadside sensors. Utility companies and infrastructure operators managing pipelines or transit corridors that run parallel to fiber routes could extend similar sensing for intrusion detection or structural monitoring, turning cable rights of way already in the ground into an added layer of situational awareness.

The fiber is already there, carrying traffic every day, so the real opportunity is in getting more value out of it without disrupting the network it was built for. What excites me about this work is that the sensing and the AI must be designed together. A method that only works in a quiet lab, or a model that only works on the one route it was trained on, will not survive contact with a live, noisy metro network.

— Ming-Fang Huang, NEC Laboratories America