基于BP神经网络的自适应偏置比例导引  

BP Neural Network-Based Adaptive Biased Proportional Navigation Guidance Law

作　　者：刘畅[1,2] 王江[1,2] 范世鹏[1,2] 李伶[3] 林德福[1,2] LIU Chang;WANG Jiang;FAN Shipeng;LI Ling;LIN Defu(School of Aerospace Engineering,Beijing Institute of Technology,Beijing 100081,China;Beijing Institute of Technology,China-UAE Belt and Road Joint Laboratory on Intelligent Unmanned Systems,Beijing 100081,China;Beijing Aerospace Automatic Control Institute,Beijing 100854,China)

机构地区：[1]北京理工大学宇航学院,北京100081 [2]北京理工大学中国-阿联酋智能无人系统“一带一路”联合实验室,北京100081 [3]北京航天自动控制研究所,北京100854

出　　处：《兵工学报》2022年第11期2798-2809,共12页Acta Armamentarii

基　　金：国家自然科学基金项目(61827901)。

摘　　要：针对传统的解析偏置比例导引在大范围机动时制导精度较差的缺陷,提出一种基于反向传播(BP)神经网络的自适应偏置比例导引律,通过BP神经网络在线精确求解偏置项。深入分析解析形式的求解偏置项的误差情况,论证不同参数对偏置项求解误差的影响。证明了参数与偏置项间存在的一一映射关系,采用BP神经网络对该映射进行高精度的拟合逼近。对多维输入参数进行灵敏度分析,以此为依据,为BP神经网络在参数空间批量化生成均衡样本。建立基于BP神经网络的偏置项求解模型,采用Adam学习方法对网络进行训练,并从理论上证明了制导律的稳定性。随后对训练效果进行测试,并对所提出的方法进行数学仿真验证。仿真结果表明:所提方法能在有限计算代价下有效提升制导精度,在不考虑弹体动力学情况下,终端角度误差均值仅为0.024°,可为工程应用提供参考。To address the drawback of traditional analytical biased proportional guidance with poor guidance accuracy when maneuvering in a wide range, an adaptive biased proportional guidance law based on BP(Back propagation) neural network is proposed. The bias term is accurately solved online through the BP neural network. Firstly, the error of solving bias term in analytic form is investigated. Specifically, the influence of different parameters on the solution error of bias term is demonstrated. Secondly, the mapping relationship between parameter and constant term is proved. BP neural network is used to fit the mapping accurately. Thirdly, sensitivity analysis was performed for multidimensional input parameters, on this basis, equilibrium samples for BP neural network in parameter space batch are generated. Finally, the bias term solution model based on BP neural network is established and Adam learning method is used to train the network. In addition, the stability of the guidance law is proved in theory. The effectiveness of the training is tested and verified by mathematical simulation. The simulation results show that the proposed method can be implemented with limited computational cost and effectively improve guidance accuracy, and the average impact angle error is 0.024°. This paper provides a reference for engineering application.

关 键 词：偏置比例导引 映射关系 灵敏度分析 BP神经网络 

分 类 号：TJ765.3[兵器科学与技术—武器系统与运用工程]