At ports as busy and large as Singapore’s, navigating a ship to safety can be fraught with challenges. With hundreds of vessels moving in and out every hour, each one must conduct a complex dance with many ever-changing partners in a confined space. Like a bustling ballroom in a historical drama, one wrong step can lead to catastrophic consequences.
To help coordinate that dance and boost maritime safety, a team of scientists from A*STAR’s Institute of High Performance Computing (IHPC) are looking to artificial intelligence (AI) technologies to help a ship’s crew and port control operator precisely anticipate moving vessels’ next steps in real time; a goal known as vessel trajectory prediction (VTP).
However, VTP is no easy feat even for robust computing systems. “Vessel movements are highly non-linear, complex and stochastic. It can also be hard to understand the intentions behind a vessel’s movements, especially in confined bodies of water,” said Xiuju Fu, an IHPC Senior Principal Scientist and Director of its Maritime AI Research Programme.
Existing VTP approaches often rely on statistical models or simulations, which are imprecise and can be difficult to interpret. “Yet without precise trajectory prediction, collision risk alerts may turn out false, or come too late for safety management,” Fu added.
Fu worked with IHPC Research Scientist Zhe Xiao and colleagues to develop DCA-MSDL, a multi-stage deep learning model with a dynamic context-aware approach. The model was developed under the Maritime AI Research Programme with support from the Singapore Maritime Institute, as well as Jimmy Koh from maritime industry partner PSA Marine.
The workflow of DCA-MSDL, comprised of four modules: port entry identification, turning status prediction, trajectory prediction and trajectory enhancement. Of these, the latter three modules are performed in real time. The model detects the vessel’s turning state every 30 seconds and carries out trajectory prediction and enhancement every one metre. The trajectory prediction algorithm ends when the vessel has passed through a designated traffic separation zone.
“DCA-MSDL constantly monitors a ship’s surrounding dynamic traffic using real-time data to create an awareness of the situation, assessing the risk metrics between itself and nearby vessels. It then adjusts its predictions of trajectories based on the vessels’ movement intentions; for example, turning into ports, or moving to pilot boarding stations,” said Xiao.
To achieve this, DCA-MSDL takes ship transponder data and processes it in four successive modules: port entry identification, turning status prediction, trajectory prediction and trajectory enhancement. Dynamic context-awareness comes into play in the final module, where the trajectory prediction is enhanced by making use of the latest available traffic data.
In comparative tests using a dataset of real-world vessel trajectories in the Singapore Strait, DCA-MSDL achieved 93.37 percent accuracy when predicting vessel turning status. The model also outperformed other state-of-the-art models and showed an improvement in prediction error by at least 33 percent. Fu and Xiao credit this to DCA-MSDL’s multi-stage design, which enables stage-by-stage optimisation and validation to boost prediction performance.
While the team’s explorations represent a significant step in VTP for safer ports, more work remains. “Before commercial use, there needs to be thorough testing and evaluation with live stream traffic data,” said Fu.
The team aims to enhance DCA-MSDL’s prediction and risk warning abilities to include likely locations and timings of vessel arrivals, which would support pilotage operations and traffic safety management.
The A*STAR-affiliated researchers contributing to this research are from the Institute of High Performance Computing (IHPC).
