Vessel tracking data just got a more rigorous anomaly detection framework — and it uses large language models to help fill the gaps human experts leave behind.
A team of researchers has published a new taxonomy for classifying maritime anomalies in Automatic Identification System (AIS) data, the GPS-like feed that ships broadcast to identify themselves and report position. The system defines three anomaly categories: unexpected AIS activity, route deviation, and close approach between vessels. On top of that taxonomy, they built a pipeline that generates synthetic anomalies, scores their plausibility with an LLM, and stamps timestamp-level labels onto the data — all without requiring a domain expert to hand-label every incident. The code is open-source and available on GitHub.
Most public AIS datasets ship with no anomaly labels at all, which has forced prior work into two bad corners: flag statistically rare behavior (which misses real hazards that happen to look normal) or pay experts to label by hand (slow, expensive, and hard to reproduce). The close-approach category is the sharpest addition — near-miss events between vessels are exactly the kind of interaction-driven hazard that purely statistical methods tend to ignore.
The benchmark includes tests across different time-window lengths and anomaly compositions, which gives future researchers a more honest comparison surface than prior work offered. Whether LLM-guided plausibility scoring holds up under adversarial or real-world conditions is still an open question — but as a starting point for a field that has been winging its labels, this is a meaningful step up.