Research data and code for: COLREGs-Constrained Dynamic Collision-Risk Assessment for Complex Ship Encounters Using AIS-Based Trajectory Prediction

This dataset supports the manuscript entitled "COLREGs-Constrained Dynamic Collision-Risk Assessment for Complex Ship Encounters Using AIS-Based Trajectory Prediction: A Reproducible Benchmark Study". The study proposes a COLREGs-constrained dynamic risk index (CC-DRI) for short-term two-vessel collision-risk assessment. The central hypothesis is that short-term trajectory prediction, conventional CPA geometry, and COLREGs rule-compliance evidence provide complementary information for identifying high-risk ship encounters and rule-relevant non-compliant responses. The repository contains a reproducible AIS-like benchmark rather than restricted operational AIS records. It includes 2,400 two-vessel encounter episodes, 148,800 one-minute AIS-like messages, a 10 min observation history, a 15 min prediction and risk-evaluation horizon, and four COLREGs-relevant encounter types: head-on, crossing-starboard, crossing-port, and overtaking. The episode-level split is 60% training, 10% validation, and 30% testing. The uploaded files include benchmark data and labels, Python scripts for benchmark generation, GRU training, CPA and CC-DRI calculation, metric evaluation, scenario-stratified analysis and rule-weight ablation, trained model checkpoint, figures, result tables, environment information, metric formulae, checksums, a data dictionary, and scenario definitions. The fixed random seed is 20260528. The code was tested with Python 3.13.5, NumPy 2.3.5, pandas 2.2.3, scikit-learn 1.8.0, Matplotlib 3.10.8, and PyTorch 2.10.0+cpu. The data are synthetic/AIS-like and contain no real vessel identities, no proprietary traffic records, and no restricted operational raw AIS data. They are provided to support reviewer verification, reproducibility, and future benchmarking of trajectory-informed and COLREGs-aware maritime collision-risk assessment methods.