欢迎访问《兵工学报》官方网站，今天是

兵工学报

• •    下一篇

基于Transformer-LSTM混合神经网络的迫弹外弹道及落点预测方法

杨守怀1，黄江流2，陈志华3，黄振贵1，吴明雨4 *，邱荣贤5，郑纯4   

  1. 1. 南京理工大学 瞬态物理全国重点实验室，江苏 南京 210094；2. 上海航天动力技术研究所，上海 201108； 3. 南方科技大学 国家卓越工程学院，广东 深圳 518055；4. 南京理工大学 能源与动力工程学院，江苏 南京 210094； 5. 陆军装备部驻南京地区军事代表局驻南京地区第四军事代表室，江苏 南京 210007
  • 收稿日期:2025-02-25 修回日期:2025-05-14
  • 基金资助:
    国家自然科学基金项目（U2341215）；中国博士后科学基金面上项目（2024M764224）

Prediction of Exterior Ballistics and Impact Point of Mortar based on Transformer-LSTM Hybrid Neural Network

YANG Shouhuai 1, HUANG Jiangliu 2, CHEN Zhihua 3, HUANG Zhengui 1, WU Mingyu 4 *, QIU Rongxian 5, ZHENG Chun 4   

  1. 1. National Key Laboratory of Transient Physics, Nanjing University of Science and Technology, Nanjing 210094, Jiangsu, China; 2. Shanghai Space Propulsion Technology Research Institute, Shanghai, 201109, China; 3. National Graduate College for Engineers, Southern University of Science and Technology, Shenzhen 518055, Guangdong, China; 4. School of Energy and Power Engineering, Nanjing University of Science and Technology, Nanjing 210094, Jiangsu, China; 5. Military Representative Bureau in Nanjing, General Armament Department of PLA, Nanjing 210007, Jiangsu, China
  • Received:2025-02-25 Revised:2025-05-14

PDF

摘要： 针对现代战场对迫弹类目标外弹道轨迹及落点的快速精准预测需求，提出基于Transformer-长短期记忆（Transformer-Long Short-Term Memory, Transformer-LSTM）混合神经网络的迫弹外弹道预测方法。利用Transformer网络提取T~T+K时刻迫弹速度与三维坐标的内在联合特征，LSTM网络将该时间序列特征作为输入，映射出T+K+1时刻的三维坐标信息。为优化网络模型，研究并分析不同滑动窗口步长对外弹道预测模型收敛性能的影响。将所提混合网络与门控循环神经网络、长短期记忆网络分别进行单步、多步和落点预测的对比分析。实验结果表明：所提混合网络对外弹道三维坐标的预测精度分别可达99.78%、99.72%、99.81%，均优于其他2个网络；所提混合网络的外弹道单步预测耗时仅为1.2 ms，大幅提升了预测精度与效率。该方法可实现精确且快速的外弹道及落点预测，为迫弹拦截任务提供更多响应时间。

关键词: 外弹道预测, 深度学习, Transformer-长短期记忆混合神经网络, 滑动窗口

Abstract: Focusing on the modern battlefield’s need for fast and accurate prediction of exterior ballistics and impact points of mortar, a mortar exterior ballistics prediction method based on Transformer-Long Short-Term Memory (Transformer-LSTM) hybrid neural network is proposed. The Transformer network is utilized to extract the intrinsic joint features of mortar’s velocity and three-dimensional coordinates at the moments from T to T+K, and the LSTM network takes these time series features as an input to map the three-dimensional coordinate information at the moment of T+K+1. In order to optimize the network model, the effects of different sliding window step sizes on the convergence performance of the exterior ballistics prediction model are investigated and analyzed. The proposed hybrid network is compared and analyzed with GRU network and LSTM network for single-step, multi-step and impact-point prediction, and it is found that the prediction accuracies of the proposed hybrid network for the three-dimensional coordinates of the exterior ballistics can reach up to 99.78 %, 99.72 %, 99.81 %, which are better than those of the other two networks; and the single-step prediction of the exterior ballistics of the proposed hybrid network consumes only 1.2 ms, which significantly improves the prediction accuracy and efficiency. The method enables accurate and fast prediction of exterior ballistics and impact point, providing more response time for mortar interception missions.

Key words: exterior ballistics prediction, deep learning, Transformer-LSTM hybrid neural network, sliding window

中图分类号: 

京公网安备11010802024360号    京ICP备05059581号-4
版权所有 © 《兵工学报》编辑部
地址：北京市海淀区车道沟10号中国兵工学会　邮编：100089
电话：010-68963060/68962718　 传真：010-68963025　 Email：bgxb@cos.org.cn
总访问量：　当前在线：