Dissertation Defence: Deep learning and information fusion for vessel destination prediction
August 20 at 9:00 am - 1:00 pm
Chengkai Zhang, supervised by Dr. Zheng Liu, will defend their dissertation titled “Deep learning and information fusion for vessel destination prediction” in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Electrical Engineering.
An abstract for Chengkai Zhang’s dissertation is included below.
Examinations are open to all members of the campus community as well as the general public. Please email zheng.liu@ubc.ca to receive the Zoom link for this exam.
Abstract
Maritime shipping underpins over 80% of global trade, making accurate vessel destination prediction essential for efficient port operations, optimized resource allocation, and strategic decision-making. However, existing methods face persistent challenges, including high computational costs for trajectory comparisons, limited effectiveness and prediction stability during early travel stages with trajectory-only signals. This thesis aims to address these gaps by developing three novel methodologies – TrajBERT Deep Semantic Similarity Model (DSSM), CrossRanker, and TrajReducer to improve prediction accuracy, robustness, and efficiency across all travel stages.
Firstly, TrajBERT-DSSM is proposed to capture spatiotemporal correlation, geometrical attributes, and vessel movement in trajectory comparisons. By applying geohash encoding to filter out redundant trajectories and employing a contextual embedding model – TrajBERT combined with DSSM, TrajBERT-DSSM effectively identifies destinations based on historical vessel paths. Experimental results demonstrate enhanced accuracy and stability, showcasing the benefits of integrating comprehensive trajectory features.
Secondly, CrossRanker addresses the limitations of single-dimensional approaches by fusing temporal, spatial, and static signals for destination prediction. Through a two-stage ranking process, CrossRanker combines trajectory-based similarity with feature-level measurements encompassing vessel type, architectural details, real-time draught, travel distance, and time. This cross-dimensional signal fusion boosts early-stage prediction accuracy and consistently outperforms state-of-the-art methods in Top-1 Accuracy and robustness metrics, reducing the uncertainty inherent in partial trajectory data.
Finally, TrajReducer tackles the computational inefficiency common in large-scale applications by clustering past trajectories according to spatial characteristics and selectively comparing them based on static and dynamic vessel metadata. This approach not only reduces the search space for trajectory comparisons but also preserves high accuracy throughout all travel stages. In evaluations, TrajReducer achieves a high reduce ratio and maintains superior accuracy even with limited trajectory signals.
Overall, the proposed methods form a comprehensive framework that advances the state of vessel destination prediction in accuracy, efficiency, and robustness. These contributions have significant implications for the maritime industry, ranging from enhanced port scheduling and traffic management to more sustainable operational strategies, thereby supporting the growing global demand for reliable maritime logistics solutions.
