基于船舶运动行为与时序图神经网络的轨迹预测研究
2024,32(10):39-46
摘要：随着我国“海洋强国”战略的提出,航运业飞速发展。海上交通量迅猛增长,碰撞事故频发,同时积累的丰富的船舶航行数据,亟需在此数据基础上,对船舶的航行位置进行长时序的预测,加强对海域的整体交通状况的认知,降低船舶碰撞事故率；为此在实验中首先对船舶自动识别系统(AIS,Automatic Identification System)数据进行预处理,剔除其中的易于去除的异常点,提出基于船舶航行特征的动态轨迹去纠缠方法去除纠缠点；其次,依据船舶航行特征提出顾及行为语义约束的时空轨迹密度自适应聚类方法对船舶运动模式进行挖掘,得到船舶典型运动行为模式。最后,针对船舶轨迹以及船舶运动模式,提出一种基于运动模式的时序图神经网络轨迹预测模型,对轨迹进行长时序预测,选取粤港澳大湾区作为实验海域,经对比实验验证,该模型在长时序预测上效果优于传统模型。
关键词：船舶自动识别系统数据;船舶轨迹预测;轨迹预处理;轨迹聚类;图神经网络;门控循环单元
Research on Trajectory Prediction Based on Ship Motion Behavior and Temporal Graph Neural Network
Abstract：With China"s promotion of the "Marine Power" strategy, the shipping industry has rapidly developed. Maritime traffic volumes have dramatically increased, leading to frequent ship collisions. There is an urgent need for long-term prediction of ship trajectories based on accumulated ship navigation data to enhance awareness of maritime traffic conditions and reduce collision rates. The study first pre-processes Automatic Identification System (AIS) data to remove easily identifiable anomalous points. It proposes a dynamic trajectory decorrelation method based on ship characteristics to remove decorrelated points. Next, a self-adaptive spatiotemporal trajectory clustering method considering behavioral constraints is proposed to mine ship motion patterns. It obtains typical ship behavior patterns. Finally, a motion pattern-based temporal graph neural network model is proposed for long-term prediction of ship trajectories and patterns. The Guangdong-Hong Kong-Macao Greater Bay Area is selected as the test region. Comparative experiments validate the proposed model outperforms traditional models in long-term prediction.
Key words：AIS data; Ship trajectory prediction; Trajectory preprocessing; Trajectory clustering; Graph neural network; Gated recurrent unit
收稿日期：2024-05-28
基金项目：中国博士后科学(2021M703021)；河北省重点研发计划项目(22340301D)；河北省博士后(B2021003031)
