End-to-end machine-learning framework for anchoring safety based on AIS trajectories

Anchoring is a low-speed yet high-risk phase of vessel operations, but its safety management has received considerably less attention than that of underway navigation. This study proposes a data-driven, end-to-end framework that transforms raw AIS (Automatic Identification System) data into real-time anchoring risk assessments. Using three years of AIS records from ten dry bulk carriers, vessel positions were processed through data cleaning, status-based segmentation, and UTM coordinate transformation. Abnormal behaviors were identified through a dual-stage unsupervised procedure that combines DBSCAN++ clustering with MLESAC geometric fitting, followed by supervised prediction using an XGBoost classifier optimized via an Upper Confidence Bound search strategy. Despite pronounced class imbalance, the model achieved a balanced accuracy of 94.1 % and an AUC of 0.99, demonstrating reliable detection performance. Model interpretation further revealed that vessel speed and deviation from the anchoring center were the most influential predictors of anomalies, confirming their dominant roles in governing anchoring stability. The study supplements existing anchoring datasets by systematically labeling and analyzing abnormal behaviors that were previously unclassified in AIS records, thereby providing a quantitative characterization of real-world anchoring dynamics. By integrating probabilistic learning with interpretable indicators, the framework enables near-real-time forecasting and early warning of dragging and drifting events. The proposed system offers a scalable, interpretable, and operationally deployable solution for anchorage safety management, supporting integration into VTS dashboards and onboard monitoring platforms. • An end-to-end framework transforms raw AIS data into anchoring risk intelligence. • Density-adaptive DBSCAN++ and MLESAC jointly enable robust anomaly detection. • Six interpretable anomaly types improve behavioral clarity and decision relevance. • UCB-guided XGBoost achieves over 94 % balanced accuracy under severe class imbalance. • A probabilistic-consequence model produces stratified anchoring risk assessments.