Transfer Learning–Based Mixture Gaussian Variational Autoencoder for Multiscenario Ship Trajectory Anomaly Detection

Abstract To enhance the generalization capability of variational autoencoder (VAE) models in ship anomaly detection and improve detection accuracy while reducing false alarms in specific water areas, this paper proposes a multiscenario ship trajectory anomaly detection algorithm based on transfer learning, applicable to inland and offshore waters. The inherently low sampling rate of automatic identification system (AIS) data in inland waterways leads to inefficient trajectory compression and sampling, high computational complexity, and slow model convergence during training. To address these challenges, ship trajectory images are first fed into an encoder–decoder architecture for feature extraction. Transfer learning is then employed to construct a Gaussian mixture variational autoencoder for trajectory image reconstruction. A discriminator is subsequently utilized to identify discrepancies between the reconstructed and original trajectories. Finally, output visualized abnormal ship trajectories and corresponding abnormal scores. Using AIS data from the Wuhan section and Tsawwassen Ferry Terminal to validate the effectiveness of the model demonstrates strong anomaly detection performance and generalization capability across both water areas. The findings contribute significantly to enhancing the intelligent monitoring level and operational efficiency of maritime authorities.