生成对抗网络小样本雷达调制信号识别算法

doi:10.19665/j.issn1001-2400.2021.06.012

摘要/Abstract

摘要：

雷达调制识别技术在电子侦察、电子支援等领域发挥着重要的作用。现有的雷达调制信号识别算法,通常采用脉内特征提取或者深度学习技术来实现。但这两种方法都存在弊端。提取脉内特征需要复杂的先验知识;深度学习技术虽然不需要复杂的先验知识,但是深度学习技术是数据驱动需要海量的数据以支撑其训练。雷达信号数据的获取又十分的困难,难以构建复杂且庞大的数据集来表征模型,因而对于深度学习技术的小样本识别方法的需求变得迫在眉睫。为此,提出增强深度卷积生成对抗网络加卷积神经网络的雷达调制信号识别算法来实现数据增强,在小样本的条件下,仍能对多种雷达调制信号实现高精度识别。经对比实验,增强深度卷积生成对抗网络加卷积神经网络的算法在信噪比为0 dB、原始样本数量为200个的条件下,较DCGAN-CNN、GAN-CNN方法识别准确率提升了约4%,较卷积神经网络方法识别准确率提升了约10%。实验结果充分验证了在小样本条件下,增强深度卷积生成对抗网络加卷积神经网络的方法能有效地提升信号识别的准确率。

关键词: 信号识别, 生成对抗网络, 小样本, 卷积神经网络, 数据增强

Abstract:

Radar modulation recognition technology plays a large part in electronic reconnaissance,electronic support and other traditional areas.Existing radar modulation signal recognition algorithms are usually implemented using intra-pulse feature extraction or deep learning techniques.Both methods have disadvantages.Extracting intra-pulse features requires complex prior knowledge.Although deep learning technology does not require complex prior knowledge,it is data-driven,which needs massive data to support its training,and it is very difficult to obtain radar signal data.It is also difficult to build a complex and huge data set to represent the model.Therefore,the need for a small-sample recognition method for deep learning technology becomes urgent.In this paper,we propose an algorithm for radar modulation signal recognition based on Enhanced Deep Convolution Generative Adversarial Networks (SDCGAN) and the Convolutional Neural Network (CNN) to achieve data enhancement.Under the condition of small samples,it is still possible to achieve high-precision identification of a variety of radar modulated signals.This paper verifies the superiority of the SDCGAN-CNN algorithm over other algorithms and the effectiveness of signal recognition under small sample conditions through comparative experiments.Under the condition of a relatively high signal-to-noise ratio,the recognition accuracy rate of other generation adversarial network and convolutional neural network methods is improved by 4%,and the recognition accuracy rate of convolutional neural network methods is increased by 10%.

Key words: signal recognition, generative adversarial network, small sample, convolutional neural network, data enhancement

中图分类号:

TN911.6

于浩洋,尹良,李书芳,吕顺. 生成对抗网络小样本雷达调制信号识别算法[J]. 西安电子科技大学学报, 2021, 48(6): 96-104.

YU Haoyang,YIN Liang,LI Shufang,LV Shun. Recognition algorithm for the little sample radar modulation signal based on the generative adversarial network[J]. Journal of Xidian University, 2021, 48(6): 96-104.

图/表 14

图1

图2

图3

图4

图5

表1

表2

图6

表3

图7

图8

图9

图10

表4

参考文献 16

[1]	陈鑫. 基于轻量化AlexNet模型的雷达信号识别算法研究[D]. 黑龙江:哈尔滨工程大学, 2019.
	CHEN Xin. Research on radar signal recognition algorithm based on lightweight AlexNet model[D]. Heilongjiang:Harbin Engineering University, 2019.
[2]	MORAITAKIS I, FARGUESM P. Feature extraction of intra-pulse modulated signals using time-frequency analysis[C]// 21st Century Military Communications.Architectures and Technologies for Information Superiority.Piscataway:IEEE, 2000:22-25.
[3]	王世强, 张登福, 毕笃彦, 等. 双谱二次特征在雷达信号识别中的应用[J]. 西安电子科技大学学报, 2012, 39(2):127-132.
	WANG Shiqiang, ZHANG Dengfu, BI Duyan, et al. The application of bispectral secondary features in radar signal recognition[J]. Journal of Xidian University, 2012, 39(2):127-132.
[4]	Vanhoy G, Schucker T, Bose T. Classification of LPI radar signals using spectral correlation and support vector machines[J]. Analog Integrated Circuits and Signal Processing, 2017, 91(2):305-313. doi: 10.1007/s10470-017-0944-0
[5]	SCHÖLKOPF B, PLATT J, HOFMANN T. Greedy Layer-Wise Training of Deep Networks[J]. Advances in Neural Information Processing Systems, 2007, 19:153-160.
[6]	HINTON G E, OSINDERO S, TEH Y W. A Fast Learning Algorithm for Deep Belief Nets[J]. Neural Computation, 2006, 18(7):1527-1554. doi: 10.1162/neco.2006.18.7.1527
[7]	LECUN Y, BENGIO Y, HINTON G. Deep learning[J]. Nature, 2015, 521(7553):436-444. doi: 10.1038/nature14539
[8]	KONG S H, KIM M, HOANG L M, et al. Automatic LPI Radar Waveform Recognition Using CNN[J]//IEEE Access.Piscataway:IEEE, 2018:4207-4219.
[9]	QU Z, MAO X, DENG Z. Radar Signal Intra-Pulse Modulation Recognition Based on Convolutional Neural Network[J]. IEEE Access, 2018, 6:43874-43884. doi: 10.1109/Access.6287639
[10]	DENG J, DONG W, SOCHER R, et al. ImageNet:a Large-Scale Hierarchical Image Database[C]// 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE, 2009:20-25
[11]	杨晓红. LPI雷达信号智能检测与识别技术[D]. 四川:电子科技大学, 2019.
	Yang Xiaohong. Intelligent detection and recognition technology of LPI radar signal[D]. Sichaun:University of Electronic Science and Technology of China, 2019.
[12]	杨懿男, 齐林海, 王红, 等. 基于生成对抗网络的小样本数据生成技术研究[J]. 电力建设, 2019, 40(5):71-77.
	YANG Yinan, QI Linhai, WANG Hong, et al. Research on small sample data generation technology based on generative confrontation network[J]. Electric Power Construction, 2019, 40(5):71-77.
[13]	陈文兵, 管正雄, 陈允杰. 基于条件生成式对抗网络的数据增强方法[J]. 计算机应用, 2018, 38(11):3305-3311.
	CHEN Wenbing, GUAN Zhengxiong, CHEN Yunjie. Data augmentation method based on conditional generative confrontation network[J]. Computer Applications, 2018, 38(11):3305-3311.
[14]	PEREZ L, WANG J. The Effectiveness of data augmentation in image classification using deep learning[C/OL]. [2020-06-10]. http://arxiv.org/abs/1712.04621.2017
[15]	GULRAJANI I, AHMED F, ARJOVSKY M, et al. Improved training of wasserstein gans[EB/OL]. [2017-3-31]https://arxiv.org/pdf/1704.00028.pdf .
[16]	MAO X D, LI Q, XIE H R, et al. Least Squares Generative Adversarial Networks[C]// Proceedings of the IEEE International Conference on Computer Vision.Piscataway:IEEE, 2017:2794-2802.

生成器			判别器
层种类	输出数据维度	激活函数	层种类	输出数据维度	激活函数
Dense1	66 536	ReLU	Conv1	512×1×16	LeakyReLU
reshape	256×2×128		Conv2	256×1×32	LeakyReLU
UpSampling1	512×2×128		zero_padding	257×2×32
Conv1	512×2×128	ReLU	Conv3	129×1×64	LeakyReLU
BatchNormalization1	512×2×128		Conv4	129×1×128	LeakyReLU
UpSampling2	1 024×2×128		Flatten	16 512
Conv2	1 024×2×64	ReLU	Desnse	1
BatchNormalization2	1 024×2×64
Conv3	1 024×2×1	tanh

层种类	输出数据维度	激活函数
Conv1	2×1 024	ReLU
Pooling2	2×512×16
Conv3	2×512×32	ReLU
Pooling4	2×256×32
Conv5	2×256×64	ReLU
Flatten^·	32 768
Dense	128	ReLU
Dense	9	Softmax

实验编号	1	2	3	4	5
生成样本数量/个	0	20	40	60	80
信号识别准确率/%	81.11	83.33	87.78	89.89	91.11

模型	SSIM	PSNR	RMSE
GAN	0.71	20.9	22.9
DCGAN	0.72	21.1	22.4
SDCGAN	0.76	23.6	17.6