2025Software & technical leadPattern mining on noisy datasets
AISTime SeriesGeospatialAnomaly DetectionSemi-/Unsupervised ML
Overview
We analyze AIS (shipboard broadcast on VHF) to surface potentially illegal or unsafe behavior without hand-labeling every case. Analysts get interpretable flags to review—grounded in geography and domain rules.
Problem
AIS is massive and noisy, often missing explicit “illegal” labels. Analysts need tools to triage credible outliers while avoiding false conclusions driven by gaps or sensor artifacts.
Solution
Engineer time-series and geospatial features, detect anomalies with semi-/unsupervised approaches, then apply rule/context checks (zones, port radii, MPAs) to reduce false positives before hand-off to humans.
Suspicious Patterns
- Dark activity: AIS silence in high-risk zones
- Loitering: slow loops/dwells over protected areas
- Rendezvous: close-proximity meetings offshore
- Route anomalies: abrupt or non-standard deviations
Methods
- Trajectory features (speed/course variance, dwell time, gaps)
- Zonal context (MPAs, EEZs, port buffers); bathymetry/weather next
- Density-based / isolation methods for anomaly discovery
- Reason codes + map playback for analyst vetting
Architecture
Ingest → clean → feature pipeline → anomaly stage → rule/context filter → review UI. A feedback loop updates thresholds and features over time.
Status & Next Steps
- Status: Working prototype; refining features and evaluation set
- Add weather/current context; enrich port call events
- Sanctions lists and known-risk vessel sets
- Labeling workflow to improve precision/recall