An AI system trained on zero labeled events still managed to find every confirmed trawler strike on a live subsea cable.
Researchers deployed a Fast-Slow Deep Support Vector Data Description detector on a real subsea cable, feeding it continuous State-of-Polarization readings — essentially measurements of how light twists as it travels through fiber. The system had no event labels to learn from. Despite that, when ranked against 122,174 recordings, all five known trawler contacts landed in the top 13 flagged anomalies. The detector also surfaced additional cable-contact events that were later independently corroborated.
Subsea cables carry roughly 95 percent of international internet traffic, and physical strikes from fishing trawlers are among the most common causes of outages. Most monitoring systems need pre-labeled incident data to train on — a scarce resource when cable faults are, fortunately, rare. A detector that works without labels sidesteps that bottleneck and could be deployed on cables where no historical incident log exists.
The results are promising, but this is a single-cable study with five confirmed events — a thin evidence base for any operator considering a full rollout. Whether the false-positive rate stays manageable on busier or noisier cable routes is the question worth asking next.