【作者】 吕伟；

【导师】 邹斌；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程（电子科学与技术）， 2025， 硕士

【摘要】 空中目标意图推理是现代防空与空域体系中的关键技术环节,对于提升空域安全管理与应对突发状况具有重要意义。传统依赖专家经验的意图推理手段难以应对不确定的环境信息,以往的意图推理方法一般是通过分析目标在某一瞬间所表现出的特征来推测其意图,忽视目标的动态属性。因此,本文借助神经网络模型,对目标在多个相邻时间节点的状态序列展开了系统性分析,从中挖掘并提取目标状态间的内在关联特征。通过对空中目标的动作加以判定和轨迹进行识别,建立样本与标签之间的映射关系,从而实现对目标意图的推理。首先,为了实现对典型空中目标进行意图推理,对典型空中目标的特性与意图关联性进行分析介绍,根据不同类型空中目标的特点,选择典型空中目标样本,分析各种意图下目标状态属性、目标轨迹和状态变化特点。在此基础上,对不同意图、不同轨迹以及不同动作下的典型空中目标时序数据进行仿真,生成的时序数据为后续的实验场景设定提供先验知识支撑,同时为算法设计的合理性与可靠性提供重要的参考。然后,基于典型空中目标特性分析之后生成的时序数据,有效整合时序运动信息和投影轨迹,实现基于时图融合的动态目标行为识别,提高了空域背景下目标行为识别的实时性与精确度。针对空中目标动作判决,本文联合时域卷积网络与添加时间模式注意力机制的Bi LSTM算法模型,该模型以典型空中目标时序数据作为输入,挖掘目标的动作变化规律和关键时序特征,提高了典型空中目标动作判决的实时性与精确度。针对空中目标轨迹识别,本文提出一种基于YOLOv8-BIFPN-EMAttention网络的空中目标轨迹识别模型,对航迹信息中的轨迹数据进行截取,将水平面投影作为输入,以全局时序视角进行轨迹识别,具有优异的精度和鲁棒性。实验结果表明,依据上述两种算法能够快速、准确地实现典型空中目标行为识别。最后,对典型空中目标的意图推理与风险等级评估方法展开研究。针对意图推理问题,本文提出一种基于KMFDT算法的模糊推理方法,利用K-Means算法对空中目标的特征数据进行聚类,确定模糊隶属度函数分段点,并通过Min-Ambiguity算法构建模糊决策树,实现目标意图的推理。针对风险等级评估问题,本文依据推理的目标意图,量化分析目标任务能力、速度、距离与航向等要素,建立全连接神经网络模型,实现空中目标风险等级的有效评估。实验结果表明,该方法能有效提高空中目标意图推理与风险等级评估的准确性。更多还原

【Abstract】 Intent recognition and inference for aerial targets is a pivotal technological component in modern air defense and airspace management systems,playing a significant role in enhancing airspace security and responding to sudden threats.Traditional methods relying on expert experience struggle to handle uncertain environmental information,as most existing intent recognition algorithms only utilize feature information from a single time point and overlook the dynamic attributes of targets.Consequently,extracting meaningful information,identifying key features,and thoroughly capturing the spatiotemporal evolution of target behavior from massive datasets has become a prominent research topic.In this context,this research introduces a neural approach to revealing the internal associations of features derived from series of target states tracked over continuous temporal steps.By performing action determination and trajectory recognition on aerial targets,we establish a mapping between samples and labels,thereby enabling robust intent recognition and inference.First,in order to enable intent recognition and inference for typical airborne targets,this study analyzes and introduces the correlation between target characteristics and potential intentions.Based on the distinct features of different airborne target types,representative samples are selected to examine their state attributes,trajectories,and state transition characteristics under various intentions.Subsequently,simulations are conducted on the time-series data of these typical airborne targets under different intentions,trajectories,and maneuvers.The generated time-series data not only provides prior knowledge to support subsequent experimental scenario configurations but also serves as a crucial reference for verifying the rationality and reliability of the proposed algorithms.Secondly,based on the time-series data generated from the analysis of typical aerial target characteristics,this work effectively integrates temporal motion information and projected trajectories to achieve dynamic target behavior recognition through spatiotemporal fusion,thereby enhancing both the real-time performance and accuracy of behavior recognition in airspace environments.For aerial target action determination,we employ a model that combines a Temporal Convolutional Network with a Bi LSTM enhanced by a temporal pattern attention mechanism.By using the time-series data of typical aerial targets as input,this model explores the changing patterns of target actions and extracts key temporal features,ultimately improving both the real-time performance and accuracy of action determination for typical aerial targets.Regarding aerial target trajectory recognition,we propose a YOLOv8-BIFPN-EMAttention-based model.This model processes trajectory data extracted from flight path information,takes horizontal-plane projections as input,and adopts a global temporal perspective for trajectory recognition,demonstrating superior accuracy and robustness.Experimental results show that,with these two algorithms,typical aerial target behaviors can be recognized quickly and accurately.Finally,research is conducted on intention reasoning and threat assessment methods for typical aerial targets.To accurately infer intentions,this paper puts forward a fuzzy logic-based reasoning strategy that leverages the K-Means Fuzzy Decision Tree model.Initially,aerial target features are clustered through the application of the K-Means clustering technique,facilitating accurate determination of segmentation points within corresponding fuzzy membership functions.Following this step,the Min-Ambiguity criterion is applied to construct a fuzzy decision tree model,enabling precise inference of target intentions.For the purpose of evaluating threats posed by aerial targets,a fully-connected neural network model has been developed.This approach first determines the intentions of the targets,subsequently integrating quantitative analyses on essential parameters,including combat effectiveness,relative velocity,spatial separation,and directional trajectory.The constructed model thereby facilitates a robust and precise evaluation of aerial threats.This model is designed to enhance the accuracy and efficiency of threat evaluation by leveraging the deep learning capabilities of fully connected network structures.Through comprehensive data processing and analysis,the model provides a reliable framework for identifying and categorizing potential threats from aerial targets.Based on experimental findings,the approaches presented in this study can significantly improve the precision with which aerial targets’intentions are inferred and their potential threats are evaluated.更多还原