一种基于迁移学习的多任务制导算法

doi:10.12382/bgxb.2023.0082

摘要/Abstract

摘要：

针对典型的飞行器制导任务,利用深度学习算法可以有效地拟合导弹飞行状态与制导指令之间的函数关系。然而当制导任务发生变化时,其二者之间的映射关系也会随之改变,从而导致在当前环境下预训练好的模型无法直接作用于新环境,重新训练制导模型需要大量的弹道数据和巨额的时间成本。为解决上述问题,基于迁移学习的思想引入域对抗神经网络,提出基于迁移学习的多任务制导算法。以1个含有大量标签数据的源域任务辅助2个含有极少量标签数据的目标域任务进行迁移学习,从而克服预训练与在线控制之间的环境差异。使用特征提取器和域判别器提取出对任务环境不敏感的关键特征,使神经网络学习到各个任务所共享的底层信息;为提高预测精度,分别设计针对不同任务的偏置加速度预测器。数值仿真结果表明:基于迁移学习的多任务制导算法实现了导弹在不同任务中的加速度指令预测。

关键词: 多约束制导, 计算制导, 深度学习, 迁移学习, 偏置比例导引

Abstract:

For typical aircraft guidance missions, the deep learning algorithm can be used to effectively fit the functional relationship between missile flight state and guidance command. However, when the guidance mission changes, the mapping relationship between them will also change. As a result, a pre-trained model in the current environment cannot directly act on a new environment, and retraining the guidance model requires a large amount of ballistic data and a huge amount of time cost. In order to solve the above problems, a domain adversarial neural network is introduced based on the idea of transfer learning, and a multitask guidance algorithm based on transfer learning is proposed. One task in the source domain containing a large amount of tag data is used to assist two tasks in the target domain containing a small amount of tag data for transfer learning, so as to overcome the environmental difference between pre-training and online control. The key features that are not sensitive to the task environment are extracted by using feature extractor and domain discriminator so that the neural network learn the underlying information shared by each task. In order to improve the prediction accuracy, the bias acceleration predictors for different tasks are designed, respectively. The simulated results show that the multitask guidance algorithm based on transfer learning can predict the acceleration instruction of a missile in different missions.

Key words: multi-constraint guidance, computational guidance, deep learning, transfer learning, biased proportional navigation

中图分类号:

V249

罗皓文, 何绍溟, 亢有为. 一种基于迁移学习的多任务制导算法[J]. 兵工学报, 2024, 45(6): 1787-1798.

LUO Haowen, HE Shaoming, KANG Youwei. A Multitask Guidance Algorithm Based on Transfer Learning[J]. Acta Armamentarii, 2024, 45(6): 1787-1798.

图/表 17

图1 描述弹目相对运动的坐标系

Fig.1 Coordinate system describing the relative motion of projectiles

表1 约束条件和优化目标

Table 1 Constraints and optimization objectives

任务	约束条件
源域任务	落角约束(-90°)	最小飞行时间
目标域任务Ⅰ	落角约束(-90°)	最大终端速度
目标域任务Ⅱ	攻击时间约束(70s)和落角约束(-90°)	最大终端速度

图2 基于迁移学习的多任务制导算法框架

Fig.2 Framework of multitask guidance algorithm based on transfer learning

图3 经典域对抗神经网络结构

Fig.3 Classical domain adversarial neural network architecture

图4 改进的域对抗神经网络

Fig.4 Improved domain adversarial neural network

图5 改进的域对抗神经网络算法

Fig.5 Improved domain adversarial neural network algorithm

表2 初始状态

Table 2 Initial state

参数	数值
导弹x方向初始位x₀/m	0
导弹y方向初始位y₀/m	5000
目标x方向位置x_f/m	[20000,21000]
目标y方向位置y_f/m	2000
导弹初始速度v₀/(m·s^-1)	300
初始弹道倾角θ₀/°	[0,10]

表3 数据集大小

Table 3 The data size of data sets

数据集	数据大小	弹道条数
源域数据	121816	440
目标域Ⅰ数据a	4642	22
目标域Ⅰ数据b	6018	33
目标域Ⅰ数据c	7407	44
目标域Ⅱ数据a	3554	22
目标域Ⅱ数据b	4506	33
目标域Ⅱ数据c	5857	44

表4 不同训练集

Table 4 Different training sets

训练集	数据组成
Ⅰ	源域数据+目标域Ⅰ数据a+目标域Ⅱ数据a
Ⅱ	源域数据+目标域Ⅰ数据b+目标域Ⅱ数据b
Ⅲ	源域数据+目标域Ⅰ数据c+目标域Ⅱ数据c

图6 TMG-Ⅰ、TMG-Ⅱ、TMG-Ⅲ在目标域Ⅰ任务的终端误差对比

Fig.6 Comparison of terminal errors of TMG-Ⅰ, TMG-Ⅱ and TMG-Ⅲ in target domain Ⅰ

图7 TMG-Ⅰ、TMG-Ⅱ、TMG-Ⅲ在目标域Ⅱ任务的终端误差对比

Fig.7 Comparison of terminal errors of TMG-Ⅰ, TMG-Ⅱ and TMG-Ⅲ in target domain Ⅱ

表5 TMG偏置加速度的预测误差

Table 5 Bias acceleration prediction error of TMG

目标域任务	均值	最大值	中位数	标准差
Ⅰ	0.18718	1.380	0.04245	0.0846
Ⅱ	0.58169	1.501	0.13951	0.0587

图8 TMG预测偏置加速度与最优解的对比

Fig.8 Comparison of TMG predicted bias acceleration with optimal solution

图9 TMG和TDM在目标域任务Ⅰ中的表现

Fig.9 Performance of TMG and TDM in Task Ⅰ in the target domain

图10 TMG和TDM在目标域任务Ⅱ中的表现

Fig.10 Performance of TMG and TDM in Task Ⅱ in the target domain

表6 目标域任务Ⅰ的误差统计

Table 6 Error statistics of Task Ⅰ in the target domain

参数	均值	最大值	中位数	标准差
Δx/m	0.00039	0.00172	0.000334	0.000342
Δy/m	0.564	0.996	0.5454	0.284
Δv/(m·s^-1)	0.0369	0.1258	0.0275	0.02983
Δθ/(°)	0.0438	0.1515	0.0343	0.0383
Δt/s	0.0323	0.114	0.029	0.0266

表7 目标域任务Ⅱ的误差统计

Table 7 Error statistics of Task Ⅱ in the target domain

参数	均值	最大值	中位数	标准差
Δx/m	0.00244	0.00685	0.00159	0.00208
Δy/m	0.6359	0.991	0.6515	0.190
Δv/(m·s^-1)	0.1251	0.3829	0.07192	0.128
Δθ/(°)	0.2221	0.4253	0.18774	0.1441
Δt/s	0.0284	0.052	0.03	0.0172

参考文献 27

[1]	肖增博, 雷虎民, 滕江川, 等. 多导弹协同制导规律研究现状及展望[J]. 航空兵器, 2011(6):18-22.
	XIAO Z B, LEI H M, TENG J C, et al. Research status and prospect of multi-missile cooperative guidance law[J]. Aeronautical Ordnance, 2011(6):18-22. (in Chinese)
[2]	张春妍, 宋建梅, 侯博, 等. 带落角和时间约束的网络化导弹协同制导律[J]. 兵工学报, 2016, 37(3): 431-438. doi: 10.3969/j.issn.1000-1093.2016.03.007
	ZHANG C Y, SONG J M, HOU B, et al. Cooperative Guidance law with impact angle and impact time constraints for networked missiles[J]. Acta Armamentarii, 2016, 37(3): 431-438. (in Chinese)
[3]	陈琦, 王中原, 常思江. 带有落角约束的间接Gauss伪谱最优制导律[J]. 兵工学报, 2015, 36(7): 1203-1212. doi: 10.3969/j.issn.1000-1093.2015.07.008
	CHEN Q, WANG Z Y, CHANG S J. Optimal guidance law with impact angle constraints based on indirect gauss pseudospectral method[J]. Acta Armamentarii, 2015, 36(7): 1203-1212. (in Chinese) doi: 10.3969/j.issn.1000-1093.2015.07.008
[4]	李强, 卢宝刚, 王晓辉, 等. 一种导弹终端侵彻多约束最优制导方法[J]. 兵工学报, 2016, 37(6): 1131-1137. doi: 10.3969/j.issn.1000-1093.2016.06.023
	LI Q, LU B G, WANG X H, et al. A terminal multi-constraint optimal guidance law for missile[J]. Acta Armamentarii, 2016, 37(6): 1131-1137. (in Chinese) doi: 10.3969/j.issn.1000-1093.2016.06.023
[5]	YOU H, ZHAO F J. Distributed synergetic guidance law for multiple missiles with angle-of-attack constraint[J]. The Aeronautical Journal, 2020, 124(1274): 533-548.
[6]	叶鹏鹏, 张蛟, 李银伢, 等. 带有不同视场约束的多导弹分布式协同制导[J]. 兵工学报, 2019, 40(3): 506-515. doi: 10.3969/j.issn.1000-1093.2019.03.008
	YE P P, ZHANG J, LI Y Y, et al. Distributed cooperative guidance of multiple missiles with heterogeneous constraints of field-of-view[J]. Acta Armamentarii, 2019, 40(3): 506-515. (in Chinese) doi: 10.3969/j.issn.1000-1093.2019.03.008
[7]	张友根, 于德海, 张友安. 基于时间最优的攻击角度控制研究[J]. 飞行力学, 2010, 28(2): 46-50.
	ZHANG Y G, YU D H, ZHANG Y A. Research on attack angle control based on time optimization[J]. Journal of Flight Mechanics, 2010, 28(2): 46-50. (in Chinese)
[8]	KIM T H, LEE C H, TAHK M J, et al. Biased PNG law for impact-time control[J]. Transactions of the Japan Society for Aeronautical and Space Sciences, 2013, 56(4): 205-214.
[9]	ARITA S, UENO S. Optimal feedback guidance for nonlinear missile model with impact time and angle constraints[C]∥Proceedings of AIAA Guidance, Navigation, and Control (GNC) Conference. Reston, VA, US: AIAA, 2013: 4785.
[10]	KIM K U, KANG S, KIM H J, et al. Realtime agile-turn guidance and control for an air-to-air missile[C]∥Proceedings of AIAA Guidance, Navigation, and Control Conference. Reston, VA, US: AIAA, 2010: 7745.
[11]	KIM B S, LEE J G, HAN H S. Biased PNG law for impact with angular constraint[J]. IEEE Transactions on Aerospace and Electronic Systems, 1998, 34(1): 277-288.
[12]	LU P. Introducing computational guidance and control[J]. Journal of Guidance, Control, and Dynamics, 2017, 40(2): 193.
[13]	SHUKLA D, LAL R, HAUPTMAN D, et al. Flight test validation of a safety-critical neural network based longitudinal controller for a fixed-wing UAS[C]∥Proceedings of AIAA Aviation 2020 Forum. Reston, VA, US: AIAA, 2020: 3093.
[14]	DWIVEDI P N, BHATTACHARYA A, PADHI R. Suboptimal midcourse guidance of interceptors for high-speed targets with alignment angle constraint[J]. Journal of Guidance, Control, and Dynamics, 2011, 34(3): 860-877.
[15]	WANG N Y, et al. Deep reinforcement learning-based impact time control guidance law with constraints on the field-of-view[J]. Aerospace Science and Technology, 2022 (128): 107765.
[16]	HE S, SHIN H S, TSOURDOS A. Computational missile guidance: a deep reinforcement learning approach[J]. Journal of Aerospace Information Systems, 2021, 18(8): 571-582.
[17]	余跃, 王宏伦. 基于深度学习的高超声速飞行器再入预测校正容错制导[J]. 兵工学报, 2020, 41(4): 656-669. doi: 10.3969/j.issn.1000-1093.2020.04.005
	YU Y, WANG H L. Deep learning-basedreentry predictor-corrector fault-tolerant guidance for hypersonic vehicles[J]. Acta Armamentarii, 2020, 41(4): 656-669. (in Chinese)
[18]	刘子超, 王江, 何绍溟. 基于深度学习的时间角度控制制导律[J]. 系统工程与电子技术, 2023, 45(11):3579-3587. doi: 10.12305/j.issn.1001-506X.2023.11.25
	LIU Z C, WANG J, HE S M. Time and angle control guidance law based on deep learning[J]. Systems Engineering and Electronics, 2023, 45(11):3579-3587. (in Chinese) doi: 10.12305/j.issn.1001-506X.2023.11.25
[19]	PAN S J, YANG Q. A survey on transfer learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2009, 22(10): 1345-1359.
[20]	GANIN Y, USTINOVA E, AJAKAN H, et al. Domain-adversarial training of neural networks[J]. The journal of machine learning research, 2016, 17(1): 2096-2030.
[21]	庞伊琼, 许华, 张悦, 等. 基于迁移元学习的调制识别算法[J]. 兵工学报, 2023, 44(10): 2954-2963. doi: 10.12382/bgxb.2022.0583
	PANG Y Q, XU H, ZHANG Y, et al. Modulation recognition algorithm based on transfer meta-learning[J]. Acta Armamentarii, 2023, 44(10): 2954-2963. (in Chinese) doi: 10.12382/bgxb.2022.0583
[22]	王育鹏, 吕帅帅, 杨宇. 基于域自适应的复合材料结构损伤识别方法[J]. 航空学报, 2022, 43(6): 526752.
	WANG Y P, LÜ S S, YANG Y, et al. Damage identificat tion method of composite structures based on domain adaptive[J]. Acta Aeronautica et Astronautica Sinica, 2022, 43(6): 526752. (in Chinese)
[23]	SHI Y, WANG Z B. A deep learning-based approach to real-time trajectory optimization for hypersonic vehicles[C]∥Proceedings of AIAA Scitech 2020 Forum. Reston, VA, US: AIAA, 2020.
[24]	SÁNCHEZ-SÁNCHEZ C, IZZO D. Real-time optimal control via deep neural networks: study on landing problems[J]. Journal of Guidance, Control, and Dynamics, 2018, 41(5): 1122-1135.
[25]	GANIN Y, LEMPITSKY V. Unsupervised domain adaptation by backpropagation[C]∥Proceedings of the 32nd International Conference on International Conference on Machine Learning. Lille, France:PMLA, 2015,37:1180-1189.
[26]	SHI Y, SONG J S, HUA Y Z, et al. Leader-Follower Formation Control for Fixed-Wing UAVs using Deep Reinforcement Learning[C]∥Proceedings of the 2022 41st Chinese Control Conference. Hefei, China: IEEE, 2022:3456-3461.
[27]	吉瑞萍, 张程祎, 梁彦, 王跃东. 基于LSTM的弹道导弹主动段轨迹预报[J]. 系统工程与电子技术, 2022, 44(6): 1968-1976. doi: 10.12305/j.issn.1001-506X.2022.06.24
	JI R P, ZHANG C Y, LIANG Y, et al. Trajectory prediction of boost-phase ballistic missile based on LSTM[J]. Systems Engineering and Electronics, 2022, 44(6): 1968-1976. (in Chinese) doi: 10.12305/j.issn.1001-506X.2022.06.24