【摘要】 高速飞行器具有飞行速度快、机动范围大、突防能力强等特点，对防御系统具有较大威胁。精准预测敌方高速飞行器在制导阶段的飞行轨迹，可以提前掌握其飞行航迹，为拦截敌方导弹提供有效的技术支持。因此，本文针对高速飞行器中制导阶段的轨迹预测问题，提出一种基于门控循环单元与KAN网络架构（Gated Recurrent Unit-Kolmogorov-Arnold Network, GRU-KAN）的轨迹预测模型。首先，建立弹道坐标系下的高速飞行器运动模型，通过纵向跳跃机动模型建立轨迹数据库。随后，利用滑动窗口对轨迹数据进行分割预处理，得到轨迹数据集。最后，基于GRU和KAN架构设计轨迹预测网络，以50 s轨迹数据为输入，输出150 s预测得到的轨迹数据。实验结果表明，该模型具有更小的网络复杂度，在经度、纬度和高度方向的最大平均预测误差分别为7.58×10^-2°、9.48×10^-3°和7.51×10¹ m,经纬度方向与传统智能时序预测模型的预测误差相差不大，但在高度方向上，预测结果相比传统的GRU预测模型提升了27.8%,相比LSTM预测模型提升了70.5%。更多还原

【Abstract】 High speed aircraft have the characteristics of fast flight speed, large maneuvering range, and strong breakthrough ability, which pose a significant threat to defense system. Accurately predicting the flight trajectory of enemy high-speed aircraft during the guidance phase can provide effective technical support for intercepting enemy missiles by mastering their flight path in advance. Therefore, this article proposes a trajectory prediction model based on Gated Recurrent Unit-Kolmogorov-Arnold Network（GRU-KAN） architecture for the guidance phase of high-speed aircraft. Firstly, establish a high-speed aircraft motion model in the ballistic coordinate system, and establish a trajectory database through a longitudinal jump maneuver model. Subsequently, the trajectory data is segmented and preprocessed using a sliding window to obtain the trajectory dataset. Finally, a trajectory prediction network is designed based on GRU and KAN architectures, with 50 s trajectory data as input and 150 s predicted trajectory data as output. The experimental results show that the model has a smaller network complexity, with maximum average prediction errors of 7.58×10^-2°, 9.48×10^-3°, and 7.51×10¹ m in the longitude, latitude, and altitude directions, respectively. The prediction errors in the longitude and latitude directions are not significantly different from those of traditional intelligent temporal prediction models, but in the altitude direction, the prediction results are 27.8% higher than traditional GRU prediction models and 70.5% higher than LSTM prediction models.更多还原