Predicting maritime accident risk using Automated Machine Learning

Machine learning (ML), particularly, Automated machine learning (AutoML) offers a range of possibilities for analysing large volume of historical maritime accident records data with advanced algorithms for integrating predictive analytics in operational and policy decision making for improving maritime safety. This study explores historical data of maritime accidents in Norwegian waters over the 40 years. The data has been utilised for analysing five major maritime accident categories: grounding, contact damage, fire or explosion, collision, and heavy weather damage. A total of 29 classification ML algorithms were trained, and the Light Gradient Boosted Trees Classifier was found best performing model with highest predictive accuracy. The three most impactful factors for accident risk are: category of navigation waters, phase of operation, and gross tonnage of vessel. Based on feature effect results, vessels sailing in narrow coastal waters, in along the way operational phase, and fishing vessels are highly vulnerable to grounding relative to other types of accidents. The results can be used as input for the entire procedure of risk analysis, from hazard identification to quantification of accident consequences, and the best performing ML algorithm can be utilized in developing a decision support system for real-time maritime accident risk assessment.