AI and Maritime Safety: Why Collision Alerts Need More Than Risk Predictions

The research offers evidence from controlled simulations, with its clearest contribution being a practical separation between the danger a system estimates and the strength of the warning it is allowed to issue.

A collision warning is only as useful as the information behind it, especially in a crowded port where several ships can influence one another's movements and a delayed position report can distort the picture. A system needs to recognise danger and judge how confidently it can present that danger to the people managing traffic.

Tymoteusz I. Miller and Irmina Durlik explore this challenge in "AI-Driven Maritime Traffic Risk Assessment for Smart Seaports: Multi-Vessel Conflict Reasoning and Runtime Safety Assurance," published in the Journal of Marine Science and Engineering. Their experimental extension of PortGuard-AIS examines interactions across groups of vessels and limits automated alert authority when the supporting data becomes unreliable.

The research offers evidence from controlled simulations, with its clearest contribution being a practical separation between the danger a system estimates and the strength of the warning it is allowed to issue.

Seeing the Traffic Around a Ship

Traditional collision checks usually look at two ships at a time, estimating how close they could get and when. Busy ports are more complicated because several ships can come together, crowd the same channel or change course, shifting the greatest danger from one pair to another.

The researchers place ships on a network map, using links to show encounters between them. Each link records the estimated risk, how reliable the information is, how the ships are moving and whether they enter a safety area around one another. The system combines this information into a risk score for the whole scene. The highest risk between two ships contributes 45%, the average of the three highest risks contributes 30%, and the connections across the group contribute 25%.

That score helps the system choose one of four states: SAFE, WATCH, WARNING or CRITICAL. A change normally needs to appear in two consecutive updates before the system accepts it, and warnings stay active until the risk drops far enough to avoid alerts repeatedly switching on and off. The method uses calculations people can examine and understand, without relying on a newly trained neural network. It supports shore-based traffic operators, leaving route choices, ship handling and navigation decisions with the people responsible.

Knowing When the Data Cannot Support a Strong Alert

Automatic Identification System, or AIS, messages help ports track ships and their movements. Late messages, missing details, inaccurate positions and gaps in coverage can make this information unreliable. The system assesses data quality separately from collision risk because poor information does not automatically mean the ships are in greater danger.

A monitoring layer checks the information and processing speed before deciding what kind of alert the system can issue. Reliable data allows normal alerts. Less reliable data limits CRITICAL alerts to WARNING and WARNING alerts to ADVISORY. HUMAN_VERIFY asks a person to assess the situation, keeping a cautious risk estimate available for review. FALLBACK stops the system from issuing an automated statement about the danger.

Individual checks help catch serious problems that an overall quality score might hide. In the study, the system limited alerts when 25% of observations were out of date and switched to fallback when that figure reached 75%. These settings would need testing and adjustment before use in a working port.

The monitoring layer identified all 18 updates containing degraded information in the original tests and did not interfere unnecessarily with normal updates. Further testing covered 134 cases involving different data problems. Of 100 previously untested combinations, 78 had seriously reduced data quality, and none was allowed to keep issuing CRITICAL automated alerts. The results show that the rules worked as expected in these tests. Testing in real ports is still needed to find out how well the system handles the problems operators face in everyday traffic.

Fewer False Alerts, With a Cost in Warning Time

Evaluation covered 16 simulated scenarios, including head-on encounters, overtaking, bottlenecks, changing conflicts and six forms of degraded AIS evidence. Thirty repeated runs shifted vessel trajectories to test whether small changes in geometry altered the results.

Under the main benchmark's predefined hazard windows, the scene-based method with temporal monitoring achieved precision of 68.9%, recall of 82.1% and an F1 score of 0.749, a measure balancing precision and recall. Its false-alert rate was 36.6%, compared with 39.4% for the pairwise method with temporal monitoring.

The gain was chiefly better alert selectivity: precision improved, recall declined slightly, and the F1 improvement was not statistically significant. False-alert episodes and their combined duration fell modestly across the scenario suite, whose deliberately difficult traffic conditions cannot represent everyday port alert rates.

A separate test defined danger through actual future separation in the simulated trajectories. At a distance threshold of 0.30 nautical miles, the pairwise temporal method scored 0.881 for F1, exceeding the scene-based method's 0.842; it performed better at the other two tested distance thresholds too. A literature-derived Vessel Conflict Ranking Operator comparator strongly outperformed the static scene score at ranking the main benchmark's hazard labels.

Requiring a risk condition to persist had the largest influence on performance. Increasing persistence from one update to three improved F1 from about 0.655 to 0.841 and reduced false alerts, cutting average warning lead time from 110 seconds to roughly 63 seconds. The two-update setting provided about 94 seconds.

The release margin reduced repeated alert switching, including from ten transitions to one in a dedicated test, at the cost of retaining some warnings longer. Other local parameter changes had relatively small effects on F1. These findings make calibration a choice involving warning speed, stability and alert burden.

What a Working Port Would Still Need to Prove

Every hazardous scenario received an early detection across the repeated runs, and the assurance layer added no delay to the first actionable output compared with the underlying scene monitor. Requests for human verification counted as actionable outputs, so preserving warning time did not mean preserving full automated authority.

Processing a scene with 40 ships and 780 ship-to-ship encounters took an average of 15.83 milliseconds. The time was 23.85 milliseconds or less for about 99% of runs, and every tested scene was processed within the study's one-second limit. Those measurements excluded live data ingestion, network transport, external sensor fusion, logging, interface rendering and competing workloads, all of which would affect performance in an operational installation.