Maritime Near-Miss prediction framework and model interpretation analysis method based on Transformer neural network model with multi-task classification variables

Abstract

The prediction and analysis of Maritime Near-Miss incidents are crucial for enhancing safety protocols and accidents. In this study, a Multi-task classification variant of the Transformer neural network model is presented, designed to predict and interpret Maritime Near-Miss data. Incident reports were collected and analyzed using maritime open source intelligence, and a multi-task model based on the Transformer neural network was developed. A framework for training structured and unstructured data to predict incident risk levels and the necessity to activate the Stop Work mechanism was built. The model incorporates BERT text classification and Multi-label synthesis minority oversampling techniques to improve feature representation and address class imbalance. Dynamic weights were used to balance the learning of the two tasks during training. Experimental results show excellent performance in both risk assessment and stop work prediction tasks. The model was interpreted using feature maps and game theory, providing a new tool for maritime safety management and offering valuable insights for risk assessment and decision-making.