Abstract-As the global economy continues to expand and technology advances, the maritime industry has become a crucial element within transportation. However, with the increase in the number of ships in waterways, maritime operations have grown more complex and riskier. The collection of ship trajectories from satellite ground joint AIS (Automatic Identification System) base station may contain non-linear and non-stationary characteristics, which can compromise maritime safety and efficiency. Therefore, to more accurately predict the spatiotemporal trajectory of ships, we need to develop models that process these non-linear and non-stationary features. This paper proposes a deep learning model for ship trajectory prediction with the Transformer as the backbone. The model leverages moving average decomposition to extract trend and residual features and applies discrete wavelet transform to filter the features of all channels. This approach enhances the model’s ability to handle non-linear and non-stationary features, providing high accuracy in ship trajectory prediction. Additionally, the channels are designed independently, and the loss function is reconstructed to ensure the model’s reliability and robustness. The model is evaluated through experiments using real AIS based ship trajectories. The results of the ablation experiment indicate that the error of the network constructed by our proposed component has been reduced by 0.277 on MAE and 0.220 on MSE.
Nonlinear and Non-stationary Feature Processing Enhanced Network with Transformer for Ship Trajectory Prediction
2023-09-22
593229 byte
Conference paper
Electronic Resource
English
Ship trajectory prediction based on transformer model
IEEE | 2022
|Ship trajectory prediction model based on transformer
SPIE | 2023
|