基于模型知识融合的图神经网络多雷达协同任务调度算法  

Multiradar Collaborative Task Scheduling Algorithm Based on Graph Neural Networks with Model Knowledge Embedding

作　　者：李浩情 余点 潘常春[1] 郁文贤[1] 李东瀛 LI Haoqing;YU Dian;PAN Changchun;YU Wenxian;LI Dongying(Shanghai Key Laboratory of Navigation and Location Based Services,School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;Paris Elite Institute of Technology,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学电子信息与电气工程学院北斗导航与位置服务上海市重点实验室,上海200240 [2]上海交通大学巴黎卓越工程师学院,上海200240

出　　处：《雷达学报(中英文)》2025年第2期470-485,共16页Journal of Radars

基　　金：上海市科学技术委员会项目(24Z511005506)。

摘　　要：现代雷达的探测、跟踪、识别等任务场景越来越复杂。任务类型的多变性,雷达资源的稀缺性和任务执行时间窗口的严格要求,使得雷达任务调度成为一类强NP-Hard问题。然而,现有的调度算法在处理涉及复杂逻辑约束的多雷达协同调度问题时适应性不足,效率不高。因此,基于人工智能(AI)的调度算法正在成为研究热点,但是AI调度算法的效率与其对问题特征的提取是否全面密切相关。如何能快速、全面地提取多雷达协同任务调度问题的共性特征,是提升这类AI调度算法效率的关键。因此,该文提出了基于模型知识融合的图神经网络(MKEGNN)调度算法。该算法首先将雷达任务协同调度问题建模为异构网络图模型,利用模型知识来优化GNN算法训练过程。算法创新在于:通过低复杂度的计算手段,获取模型的关键知识,进而优化GNN模型。在特征提取阶段,引入随机酉矩阵变换,利用任务异构图的随机拉普拉斯矩阵谱特征作为全局特征来强化图神经网络对共性特征的提取能力,弱化特定问题的个性化特征;在参数化决策阶段,利用由问题的引导解和经验解构成的上/下界结构知识从原理上减少决策空间大小,引导网络快速优化,加速决策学习过程的收敛。最后,进行了大量数据仿真实验。结果表明,相比目前的算法,MKEGNN算法对于所有任务集在稳定性和精度方面都有所提升,调度成功率性能提升3%~10%,加权调度成功率提升5%~15%。尤其当处理多雷达协同关系复杂的任务集时,任务调度成功率提升4%以上,算法稳定性和鲁棒性显著增强。Modern radar systems face increasingly complex identification.The diversity of task types,limited data resources,and strict execution time requirements make radar task scheduling a strongly NP-hard problem.However,existing scheduling algorithms struggle to efficiently handle multiradar collaborative tasks involving complex logical constraints.Therefore,Artificial Intelligence(AI)-based scheduling algorithms have gained significant attention.However,their efficiency is heavily dependent on effectively extracting the key features of the problem.The ability to quickly and comprehensively extract common features of multiradar scheduling problems is essential for improving the efficiency of such AI scheduling algorithms.Therefore,this paper proposes a Model Knowledge Embedded Graph Neural Network(MKEGNN)scheduling algorithm.This method frames the radar task collaborative scheduling problem as a heterogeneous network graph,leveraging model knowledge to optimize the training process of the Graph Neural Network(GNN)algorithm.A key innovation of this algorithm is its capability to capture critical model knowledge using low-complexity calculations,which helps to further optimize the GNN model.During the feature extraction stage,the algorithm employs a random unitary matrix transformation.This approach utilizes the spectral features of the random Laplacian matrix from the task’s heterogeneous graph as global features,enhancing the GNN’s ability to extract shared problem features while downplaying individual characteristics.In the parameterized decision-making stage,the algorithm leverages the upper and lower bound knowledge derived from guiding and empirical solutions of the problem model.This strategy significantly reduces the decision space,enabling the network to optimize quickly and accelerating the learning process.Extensive simulation experiments confirm the effectiveness of the MKEGNN algorithm.Compared to existing approaches,it demonstrates improved stability and accuracy across all task sets,boosting the schedu

关 键 词：雷达任务调度 图神经网络 强化学习 模型知识 拉普拉斯矩阵 随机矩阵 

分 类 号：TN958.92[电子电信—信号与信息处理] TP389.1[电子电信—信息与通信工程]