Detecting abnormal ship trajectory to avoid bridge collisions via a Transformer-BiLSTM model

The risk of ship-bridge collision is increasing worldwide due to the growth of ship size and vessel traffic density in inland waterways and port areas. With the imminent demand for advanced real-time risk evaluation tools, machine learning approaches able to deal with big data come into play to investigate ship trajectory anomalies. This paper presents a novel data-driven approach for accurate abnormal ship trajectory detection based on the Transformer-BiLSTM reconstruction machine learning model. Ship trajectories in the vicinity of a bridge area are obtained from Automatic Identification System (AIS) data. These are then used to analyse the ships' behaviour and to formulate normal ship traffic patterns. After establishing a benchmark scenario, an unsupervised model is developed to detect the ship's abnormal behaviour using a Transformer-BiLSTM encoder-decoder architecture. The encoder extracts the traffic features, and the decoder reconstructs the ship trajectory. The abnormal behaviour is identified based on significant reconstruction errors, which indicate deviations from expected navigational patterns. A case study is conducted near the Sutong Bridge in the Yangtze River estuary in China. The model is trained on 664,517 trajectories involving 8137 ships. The model detected 33 ships with abnormal trajectories in a validation dataset of 400 ships navigating the bridge water area. Compared with other machine learning and deep learning models, the proposed approach achieved the highest accuracy of 97.04 %. The approach can substantially benefit the maritime safety authorities in effectively identifying potential risks from short segments of historical ship trajectories, enhancing early detection of hazardous navigation behaviours. • A Data-driven approach for abnormal ship trajectory detection is proposed. • The Transformer-BiLSTM machine learning model is used. • The abnormal behaviour is identified based on significant reconstruction errors. • The proposed approach achieved the highest accuracy.