一种利用贝叶斯优化的弹道目标微动分类网络

doi:10.19665/j.issn1001-2400.2021.05.017

摘要/Abstract

摘要：

弹道目标识别在现在的军事反导系统中具有重要作用。不同的弹道目标由于其运动特征的不同会呈现出不同的微动特征,因此在弹道目标的识别中,微动特征被广泛应用。在弹道目标的微动特征分类中,卷积神经网络因为其优异的分类能力而被广泛应用。在卷积神经网络框架中,神经网络相关的超参数和结构通常与被处理的任务性质相关,不同的任务需要设置不同的超参数和不同的网络架构使卷积神经网络的识别效果达到最优。人工选择适合弹道目标微动分类的网络超参数和结构需要足够的人力经验和大量的计算时间,而且并不总能得到最优的参数。针对这个问题,提出用贝叶斯优化算法来自动获取应用于弹道目标微动分类的卷积神经网络的超参数和最优结构的方法,以提高神经网络对微动特征的分类性能。仿真结果表明,与基于迁移学习的神经网络和传统的方法相比,贝叶斯优化算法得到的卷积神经网络模型具有良好的识别效果和鲁棒性,能够精准地对不同微动特征下的信号进行分类与识别。

关键词: 目标识别, 微多普勒, 贝叶斯优化, 深度学习

Abstract:

Ballistic target recognition plays an important role in the current military anti-missile system.Different ballistic targets show different fretting characteristics due to their different motion characteristics,so fretting features are widely used in ballistic target recognition.Since artificial selection for ballistic target classification of the micro structure of neural network parameters needs human experience and a large amount of computing time,and does not guarantee the optimal parameters of the problem,we suggest using the bayesian optimization algorithm to automatically obtainthe convolution method for neural network parameters and the optimal structure,in order to improve the classification performance of the neural network for the micro motion features.Experimental results show that the bayesian optimization algorithm can quickly accomplish the parameter optimization of the convolutional neural network,and that the convolutional neural network model has a good recognition effect and is robust.

Key words: target recognition, micro-doppler, Bayesian optimization, deep learning

中图分类号:

TN957

李鹏,冯存前,许旭光,唐子翔. 一种利用贝叶斯优化的弹道目标微动分类网络[J]. 西安电子科技大学学报, 2021, 48(5): 139-148.

LI Peng,FENG Cunqian,XU Xuguang,TANG Zixiang. Ballistic target fretting classification network based on Bayesian optimization[J]. Journal of Xidian University, 2021, 48(5): 139-148.

图/表 14

图1

表1

图2

图3

图4

表2

图5

表3

图6

图7

图8

表4

图9

表5

参考文献 19

[1]	CHEN V C. Analysis of Radar Micro-Doppler with Time-Frequency Transform[C]// Proceedings of the 10th IEEE Workshop on Statistical Signal and Array Processing.Piscataway:IEEE, 2000:463-466.
[2]	马梁. 弹道中段目标微动特性及综合识别方法[D]. 长沙:国防科学技术大学, 2011.
[3]	陈小龙, 关键, 何友. 微多普勒理论在海面目标检测中的应用及展望[J]. 雷达学报, 2013, 2(1):123-134.
	CHEN Xiaolong, GUAN Jian, HE You. Application and Prospect of Micro-Motion Theory in the Detection of Sea Surface Target[J]. Journal of Radars, 2013, 2(1):123-134.
[4]	韩勋, 杜兰, 刘宏伟, 等. 基于时频分布的空间锥体目标微动形式分类[J]. 系统工程与电子技术, 2013, 35(4):684-691.
	HAN Xun, DU Lan, LIU Hongwei, et al. Classification of Micro-Motion Form of Space Cone-Shaped Objects Based on Time-Frequency Distribution[J]. System Engineering and Electronics, 2013, 35(4):684-691.
[5]	关永胜, 左群声, 刘宏伟, 等. 空间锥体目标微动特性分析与识别方法[J]. 西安电子科技大学学报, 2011, 38(2):105-111.
	GUAN Yongsheng, ZUO Qunsheng, LIU Hongwei, et al. Micro-Motion Characteristic Analysis and Recognition of Cone-Shaped Targets[J]. Journal of Xidian University, 2011, 38(2):105-111.
[6]	束长勇, 张生俊, 黄沛霖, 等. 基于微多普勒的空间锥体目标微动分类[J]. 北京航空航天大学学报, 2017, 43(7):1387-1394.
	SHU Changyong, ZHANG Shengjun, HUANG Peilin, et al. Micro-Motion Classification of Spatial Cone Target Based on Micro-Doppler[J]. Journal of Beijing University of Aeronautics and Astronsutics, 2017, 43(7):1387-1394.
[7]	SINGH A, KHAN M A, BAGHEL N. Face Emotion Identification by Fusing Neural Network and Texture Features:Facial Expression[C]// 2020 International Conference on Contemporary Computing and Applications.Piscataway:IEEE, 2020:187-190.
[8]	张儒鹏, 于亚新, 张康, 等. 基于OI-LSTM神经网络结构的人类动作识别模型研究[J]. 计算机科学与探索, 2018, 12(12):1926-1939.
	ZHANG Rupeng, YU Yaxin, ZHANG Kang, et al. Research on Human Action Recognition Model Based on OI-LSTM Neural Network Structure[J]. Journal of Frontiers of Computer Science and Technology, 2018, 12(12):1926-1939.
[9]	衣世东. 基于深度学习的图像识别算法研究[J]. 网络安全技术与应用, 2018, 1:39-41.
	YI Shidong. Research on Image Recognition Algorithm Based on Deep Learning[J]. Network Security Technology and Application, 2018, 1:39-41.
[10]	LI C, LAN H Q, SUN Y N, et al. Detection Algorithm of Defects on Polyethylene Gas Pipe Using Image Recognition[J]. International Journal of Pressure Vessels and Piping, 2021, 191:104381. doi: 10.1016/j.ijpvp.2021.104381
[11]	向前, 王晓丹, 李睿, 等. 基于DCNN的弹道中段目标HRRP图像识别[J]. 系统工程与电子技术, 2020, 42(11):2426-2433.
	XIANG Qian, WANG Xiaodan, LI Rui, et al. HRRP Image Recognition of Midcourse Ballistic Targets Based on DCNN[J]. Systems Engineering and Electronics, 2020, 42(11):2426-2433.
[12]	李江, 冯存前, 王义哲, 等. 基于深度卷积神经网络的弹道目标微动分类[J]. 空军工程大学学报:自然科学版, 2019, 20(4):97-104.
	LI Jiang, FENG Cunqian, WANG Yizhe, et al. Micro-Motion Classification of Ballistic Targets Based on Deep Convolutional Neural Network[J]. Journal of Air Force Engineering University:Natural Science Edition, 2019, 20(4):97-104.
[13]	HAN L, FENG C. Micro-Doppler-Based Space Target Recognition with a One-Dimensional Parallel Network[J]. International Journal of Antennas and Propagation, 2020, 2020:1-10.
[14]	OBAYYA M I, EL-GHANDOUR M, ALROWAIS F. Contactless Palm Vein Authentication Using Deep Learning With Bayesian Optimization[J]. IEEE Access, 2021, 9:1940-1957. doi: 10.1109/Access.6287639
[15]	BERGSTRA J, BARDENET R, BENGIO Y, et al. Algorithms for Hyper-Parameter Optimization[C]// Advances in Neural Information Processing Systems.San Francisco:Morgan Kaufmann, 2011:1-9.
[16]	SNOEK J, LAROCHELLE H, ADAMS R P. Practical Bayesian Optimization of Machine Learning Algorithms[J]. Advances in Neural Information Processing Systems, 2012, 4:1-9.
[17]	RASMUSSEN C E, WILLIAMS C K I. Gaussian Processes for Machine Learning[M]. Cambridge: The MIT Press, 2006:105-128.
[18]	MURPHY K P. Machine Learning:A Probabilistic Perspective[M]. Cambridge: The MIT Press, 2012:36-54.
[19]	BELKINAA C, CICCOLELLA C O, ANNO R, et al. Automated Optimized Parameters for T-Distributed Stochastic Neighbor Embedding Improve Visualization and Analysis of Large Datasets[J]. Nature Communications, 2019, 10:1-12. doi: 10.1038/s41467-018-07882-8

微动参数	自旋	进动	章动
α/(°)	95:5:150	95:5:150	95:5:150
θ₀/(°)	12:2:20	12:2:20	12:2:20
φ₀/(°)	18:18:360	0	0
f_φ/Hz		0.82:0.02:1.2	0.84:0.04:1.2
θ₁/(°)			10
f_θ/Hz			1:0.5:15

条件	综合识别率
条件	-15 dB	-10 dB	-5 dB	0 dB	5 dB	10 dB	15 dB
有贝叶斯优化的CNN网络	60.1	72.3	91.4	97.3	100.0	100.0	100.0
无贝叶斯优化的CNN网络	56.3	68.5	84.1	92.7	97.4	98.4	99.5

参数	AlexNet	GoogleNet	SqueezeNet	文中网络
识别时间/s	2.65	7.83	3.61	3.87
帧数/(张·秒^-1)	377	128	277	258