航空学报 > 2023, Vol. 44 Issue (14): 327964-327964   doi: 10.7527/S1000-6893.2022.27964

基于改进二阶优化器并行学习的弹道导弹神经网络落点预测方法

任乐亮1(), 鲜勇1, 李少朋1,2, 雷刚1, 伍薇1, 李冰1   

  1. 1.火箭军工程大学 作战保障学院,西安 710025
    2.清华大学 自动化系,北京 100084
  • 收稿日期:2022-09-01 修回日期:2022-09-16 接受日期:2022-11-05 出版日期:2023-07-25 发布日期:2022-11-17
  • 通讯作者: 任乐亮 E-mail:renleliang@126.com
  • 基金资助:
    国家自然科学基金(62103432);陕西省高校科协青年人才托举计划(20210108)

A neural network model for impact point prediction of ballistic missile based on improved second-order optimizer with parallel learning

Leliang REN1(), Yong XIAN1, Shaopeng LI1,2, Gang LEI1, Wei WU1, Bing LI1   

  1. 1.College of War Support,Rocket Force University of Engineering,Xi’an 710025,China
    2.Department of Automation,Tsinghua University,Beijing 100084,China
  • Received:2022-09-01 Revised:2022-09-16 Accepted:2022-11-05 Online:2023-07-25 Published:2022-11-17
  • Contact: Leliang REN E-mail:renleliang@126.com
  • Supported by:
    National Natural Science Foundation of China(62103432);Young Talent fund of University Association for Science and Technology in Shaanxi(20210108)

摘要:

针对弹道导弹大机动突防后精确制导面临的落点预测需求,提出了一种基于改进二阶优化器学习的神经网络落点预测方法。基于椭圆弹道理论对当前飞行状态的落点进行预测,再求解与真实落点的偏差,并对偏差量进行解耦处理,进而构建了以飞行状态量为输入、以偏差量为输出的样本集,大幅降低了神经网络学习难度。为提高神经网络预测精度,采用3个神经网络分别预测偏差量的3个分量;利用矩阵分块运算法则建立了适用于多GPU并行的改进Levenberg-Marquardt优化器,缩短了网络学习时间且降低了对GPU显存的需求量。设计了详细的仿真实验对该方法的优势和计算复杂度进行了分析,仿真结果表明,落点预测模型的学习难度小,预测精度高,实时性好。在训练集和测试集所含869 320个样本中,3σ预测误差为4.97 m;在含2块GPU的学习环境中,学习耗时缩短约49.18%;在STM32F407单片机上,落点预测耗时为2.585 ms,能够为制导算法设计提供支撑,具有一定的工程应用价值。

关键词: 弹道导弹, 大机动突防, 落点预测, 神经网络, 并行学习, 椭圆弹道, 单片机

Abstract:

To address the requirement for Impact Point Prediction (IPP) in precision guidance for the ballistic missile after high maneuver penetration, an IPP Neural Network (NN) model is proposed based on an improved second-order optimizer. The IPP of the current flight state is predicted based on the elliptical trajectory theory, and the impact deviation is decoupled. Furthermore, a sample set with flight state as input and impact deviation of elliptical trajectory as output is constructed, which greatly reduces the difficulty of NN learning. To improve the prediction accuracy, three NNs are applied to predict the three components of the impact deviation. An improved Levenberg-Marquardt optimizer for multi-GPU parallel learning is established by using the matrix block algorithm, which shortens the network learning time and reduces the demand for GPU memory. A simulation experiment is designed to analyze the advantages and computational complexity of the proposed method. Simulation results show that the proposed method has low learning difficulty, high prediction accuracy and good real-time performance. Among the 869 320 samples contained in the training set and the test set, the 3σ prediction error is 4.97 m. In the learning environment with 2 GPUs, the learning time is reduced by about 49.18%. On the STM32F407 single-chip microcomputer, IPP takes 2.585 ms. The proposed method can provide support for the design of guidance algorithm, and is applicable for engineering practice.

Key words: ballistic missile, high maneuver penetration, impact point prediction, neural network, parallel learning, elliptical trajectory, single-chip microcomputer

中图分类号: