基于时间感知图卷积网络的船舶航迹预测  

Ship trajectory prediction based on time-aware graph convolutional network

作　　者：王宇 侯凌燕[1] 王超[1] 赵青娟 邹智元 WANG Yu;HOU Lingyan;WANG Chao;ZHAO Qingjuan;ZOU Zhiyuan(Computer School,Beijing Information Science&Technology University,Beijing 102206,China)

机构地区：[1]北京信息科技大学计算机学院,北京102206

出　　处：《北京信息科技大学学报（自然科学版）》2024年第4期41-48,共8页Journal of Beijing Information Science and Technology University

基　　金：国家重点研发计划项目(2022YFF0604502)。

摘　　要：针对船舶航迹数据非等距时间间隔的特点,提出一种基于时间感知图卷积网络的船舶航迹预测方法。该方法将船舶的时序航迹数据转换为图结构,结合改进的高斯核函数,以图节点间的时间距离和空间欧式距离构建时空带权邻接矩阵,显式建模非等距时间间隔采样点间的时空关联强度,以此捕捉船舶航迹数据的时空动态关系;通过图卷积网络和自注意力机制,依据时空带权邻接矩阵,提取图中节点的时空特征并加权,使模型聚焦于船舶航迹的关键时空信息。在船舶自动识别系统采集的公开数据上进行的仿真实验结果显示,与基于循环神经网络的方法相比,本文方法预测精度有显著提升。According to the characteristics of irregular time intervals of ship track data,a ship trajectory prediction method based on time-aware graph convolutional network(T-GCN)was proposed.This method converted the time-series trajectory data of ships into a graph structure,and combined an improved Gaussian kernel function to construct a spatiotemporal weighted adjacency matrix based on the time distance and spatial Euclidean distance between graph nodes.It explicitly modeled the spatiotemporal correlation strength between non equidistant time interval sampling points in order to capture the spatiotemporal dynamic relationship of ship trajectory data.By using graph convolutional networks and self-attention mechanisms,the spatiotemporal features of nodes in the graph were extracted and weighted based on the spatiotemporal weighted adjacency matrix,enabling the model to focus on key spatiotemporal information of ship trajectories.Simulation experiments were conducted on publicly available data collected by the ship automatic identification system(AIS),and the results show that compared with the method based on recurrent neural networks,the proposed method significantly improves prediction accuracy.

关 键 词：航迹预测 非等距时间间隔 带权邻接矩阵 图卷积网络 自注意力机制 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]