关键词: 船舶交通工程；船舶类型识别；AIS轨迹数据；图同构网络；注意力机制；类别不平衡

Abstract: To address the problem that existing graph neural networks diluted critical features of long trajectories and failed to distinguish fine-grained category differences due to average or sum pooling in the graph readout phase, a graph isomorphism network fusing a global attention mechanism (GIN-GA) was proposed. The improvements of the proposed model included: retaining the Delaunay triangulation spatiotemporal graph composition strategy to preserve topological integrity; introducing global attention pooling in the feature aggregation layer to adaptively assign trajectory point weights, and enabling the model to focus on the most discriminative key nodes for classification; employing a weighted cross-entropy loss to address the class imbalance problem in AIS data. Research results show that GIN-GA achieves an F1 score of 92.81%, which is 3.32%, 2.92%, 3.01%, and 2.29% higher than those of the original GIN, mainstream graph convolutional networks (GCN), graph attention networks (GAT), and graph sample and aggregate networks (GraphSAGE), respectively. The recall rates of two minority samples, passenger ships and tugboats, are improved by 4.70% and 7.00%, respectively. The proposed model effectively enhances ship recognition accuracy and provides a reference for improving the intelligence level of maritime supervision.

Key words: ship traffic engineering; ship type recognition; AIS trajectory data; graph isomorphism network; attention mechanism; class imbalance