Ship Collision Risk Evaluation using AIS and weather data through fuzzy logic and deep learning

According to the European Maritime Safety Agency, marine casualties and incidents are mainly caused by 'Human action' which highlights the need for more advanced technologies to enhance the safety of ships, particularly through improved risk assessment of ship collisions. Accordingly, this study explores the application of Deep Learning for Ship Collision Risk Evaluation to detect ship collisions. A Collision Risk Index (CRI) between two vessels at a given instance is calculated using AIS data. The environmental impact is captured using an Environmental Factor (Enfactor) index for own and target ships. These indices are combined using different weights to obtain a Resultant that represents collective impact of weather factors on both the ships. Subsequently, CRI and Resultant Enfactor indices are combined using Fuzzy Logic to obtain a comprehensive index called Enhanced Collision Risk Index with Weather (ECRI-W). Various deep learning models are evaluated and the Self-Attention and Intersample Attention Transformer (SAINT) model is selected due to its superior performance with tabular data. The Piraeus AIS and Weather datasets are used for model training and testing. After preprocessing the datasets, the model undergoes self-supervised pretraining and supervised finetuning. It is concluded that SAINT-s variant with pretraining achieves the best performance.