TrAISformer -- A Transformer Network with Sparse Augmented Data Representation and Cross Entropy Loss for AIS-based Vessel Trajectory Prediction

Duong Nguyen,Ronan Fablet

DOI: https://doi.org/10.1109/ACCESS.2024.3349957

2024-01-03

Abstract:Vessel trajectory prediction plays a pivotal role in numerous maritime applications and services. While the Automatic Identification System (AIS) offers a rich source of information to address this task, forecasting vessel trajectory using AIS data remains challenging, even for modern machine learning techniques, because of the inherent heterogeneous and multimodal nature of motion data. In this paper, we propose a novel approach to tackle these challenges. We introduce a discrete, high-dimensional representation of AIS data and a new loss function designed to explicitly address heterogeneity and multimodality. The proposed model-referred to as TrAISformer-is a modified transformer network that extracts long-term temporal patterns in AIS vessel trajectories in the proposed enriched space to forecast the positions of vessels several hours ahead. We report experimental results on real, publicly available AIS data. TrAISformer significantly outperforms state-of-the-art methods, with an average prediction performance below 10 nautical miles up to ~10 hours.

Artificial Intelligence

What problem does this paper attempt to address?

The paper aims to address the problem of ship trajectory prediction in Automatic Identification System (AIS) data, particularly the challenges of medium to long-term predictions (ranging from tens of minutes to several hours).

The paper points out that although AIS provides a rich source of information to tackle the ship trajectory prediction task, the inherent heterogeneity and multimodal nature of motion data make it difficult for even modern machine learning techniques to accurately predict ship trajectories. To overcome these challenges, the authors propose a new method called TrAISformer. This method mainly includes the following aspects:

1. **Novel Data Representation**: To better handle the heterogeneity of AIS data, the authors introduce a discrete and high-dimensional data representation method, transforming continuous observations into sparse "four-hot" vectors, which are further mapped into high-dimensional embedding vectors.

2. **Classification Loss Function**: Considering the potential multimodal distribution in ship trajectory prediction, the authors use a cross-entropy loss function for training to explicitly model the multimodality of the data.

3. **Improved Transformer Architecture**: By leveraging the powerful capabilities of the Transformer network to capture long-term dependencies in AIS ship trajectories, the prediction accuracy is improved.

Experimental results show that TrAISformer achieves significantly better performance than existing methods on real publicly available AIS datasets, with an average prediction error of less than 10 nautical miles and a prediction time range of up to approximately 10 hours. Additionally, the paper provides detailed technical details, experimental settings, and result analysis of the proposed method, validating its effectiveness and superiority.