Abnormal behaviour detection methods in normal navigation modes of ships

To address the challenges of characterising normal ship behaviours in complex maritime scenarios, scarce abnormal samples, and limited interpretability, this paper proposes an unsupervised abnormal ship behaviour detection method based on normal navigation modes using AIS data. First, features including speed, speed change rate, heading change rate, and position change rate are selected to construct a normal behaviour scoring model, and normal samples are screened via an adaptive threshold. Subsequently, a ship behaviour prediction model is established to predict future navigation states, extending anomaly detection from historical-state identification to early perception of future risks. A variational autoencoder then learns the distribution characteristics of normal samples, and an abnormal score is constructed using reconstruction errors to recognise abnormal behaviours. Additionally, the influence of different features on abnormal results is analysed to explain the causes of anomalies. Experimental results based on AIS data from the Miami waters, California coastal waters, and northern waters of Hangzhou demonstrate that the proposed method effectively identifies abnormal behaviours across different regions, with F1-scores of 0.6250, 0.6957, and 0.6667 in the prediction domains, respectively, indicating good scene adaptability.