战场目标实体类型识别的鲁棒图神经网络方法  

Robust graph neural network method for target entity type recognition in a battlefield environment

作　　者：周贤琛 马扬 程光权[2] 王红霞[1] ZHOU Xianchen;MA Yang;CHENG Guangquan;WANG Hongxia(College of Liberal Arts and Sciences,National University of Defense Technology,Changsha 410072,China;College of Systems Engineering,National University of Defense Technology,Changsha 410072,China)

机构地区：[1]国防科技大学文理学院,湖南长沙410072 [2]国防科技大学系统工程学院,湖南长沙410072

出　　处：《智能系统学报》2023年第6期1156-1164,共9页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61977065,62073333).

摘　　要：随着信息战和算法战等新型作战样式的兴起,战场数据分析中的目标实体识别任务对决策起着重要作用。战场态势数据作为最典型的战场数据之一,包含了许多紧密交互的动态实体数据。但此类数据因敌方干扰或伪装常常含有较强的噪声,与一般时序关系数据相比,对目标实体方法的鲁棒性要求更高。本文基于图神经网络提出了一种表示和处理这类态势数据、识别敌方作战实体的新方法。首先,使用动态时间规整算法基于作战实体轨迹建立了作战实体之间的新型图结构数据模型,进而根据作战实体的节点属性信息提出了一种鲁棒的图神经网络方法,并将其应用于雷达识别范围之外的作战实体类型辨识。在兵棋推演平台获得的仿真数据集上的测试结果表明,本文方法由于充分利用了实体数据的时序特征以及关联的属性信息,与依赖单个时刻关系构建出的图神经网络方法以及多层感知机等方法相比,在识别精度、鲁棒性等方面优势明显,一定程度上扩大了作战实体识别的半径。With the rise of new combat styles,such as information and algorithmic warfare,target entity recognition in battlefield data analysis plays an important role in decision making.Battlefield situation data are typical battlefield data containing many dynamic entities with close interactions.However,such data often contain strong noise due to hostile interference or concealment;hence,they require higher robustness than general time-series data.This paper proposes a new method based on graph neural networks to represent and process the unstructured data and mine the category information of hostile combat entities.First,the dynamic time warping algorithm was used to establish a new graph structure between combat entities based on their trajectory.Then,a robust graph neural network method was proposed and applied for the type identification of combat entities beyond the radar identification range according to the node attribute information of combat entities.Test results on the simulation data set obtained from the military simulation platform reveal that the proposed method maximizes the temporal characteristics of the entity data and associated attribute information of each node.Compared with the graph neural network and multilayer perceptron methods that rely on singletime relation,the proposed method has advantages in identification accuracy and robustness,expanding the radius of operational entity identification to a certain extent.

关 键 词：战场态势数据 实体识别 识别半径 动态时间规整 数据挖掘 图神经网络 鲁棒性 图卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]