基于特征融合的Fast-CNN的复杂波形调制识别

doi:10.12305/j.issn.1001-506X.2025.07.10

摘要/Abstract

摘要：

针对复杂电磁环境中, 雷达信号种类复杂、形式敏捷多变, 导致信号识别率较低与运算复杂度较高的问题, 提出一种基于特征融合的快速卷积神经网络(fast-convolutional neural network, Fast-CNN)的复杂波形调制识别方法。首先, 设计规模较大的教师网络(特征融合网络)和规模较小的学生网络(Fast-CNN)。教师网络提取并融合特征图的不同尺度特征, 提高网络识别率; 学生网络通过剪枝方法去除冗余通道, 解决计算量较大的问题。然后, 通过知识蒸馏将教师网络训练得来的知识转移到学生网络中, 从而网络能在显著降低计算量的同时保持识别精度。实验表明, 当信噪比大于-3 dB时, 所提方法对10类复杂调制波形的整体识别率达到99%以上。

关键词: 复杂波形, 调制识别, 特征融合, 网络剪枝, 知识蒸馏

Abstract:

Aiming at the complex electromagnetic environment, the radar signals are complex in type and agile in form, which leads to the problem of lower signal recognition rate and higher arithmetic complexity, a complex waveform modulation recognition method based on fast-convolutional neural network (Fast-CNN) with feature fusion is proposed. Firstly, a larger teacher network (feature fusion network) and a smaller student network (Fast-CNN) are designed. The teacher network extracts and fuses the different scale features of the feature map to improve the network recognition rate. The student network removes the redundant channels by pruning method to solve the problem of larger computation. Then, the knowledge trained in the teacher network is transferred to the student network through knowledge distillation. Thus, the network can significantly reduce the amount of computation while maintaining the recognition accuracy. Experiments show that the overall recognition rate of the proposed method for 10 types of complex modulated waveforms reaches more than 99% when the signal-to-noise ratio is greater than -3 dB.

Key words: complex waveform, modulation recognition, feature fusion, network pruning, knowledge distillation

中图分类号:

TN974

周东平, 阮航, 沙明辉, 王崇宇, 崔念强, 鲁耀兵. 基于特征融合的Fast-CNN的复杂波形调制识别[J]. 系统工程与电子技术, 2025, 47(7): 2165-2175.

Dongping ZHOU, Hang RUAN, Minghui SHA, Chongyu WANG, Nianqiang CUI, Yaobing LU. Complex waveform modulation recognition based on feature fusion with Fast-CNN[J]. Systems Engineering and Electronics, 2025, 47(7): 2165-2175.

图/表 21

图1

图2

图3

图4

图5

表1

图6

图7

图8

图9

图10

图11

表2

图12

图13

表3

图14

表4

表5

表6

图15

参考文献 41

1	欧健. 多功能雷达行为辨识与预测技术研究[D]. 长沙: 国防科技大学, 2017.
	OU J. Research on behavior recognition and prediction techniques against multi-function radar[D]. Changsha: National University of Defense Technology, 2017.
2	薛坚. 雷达辐射源信号识别关键技术研究[D]. 南京: 南京大学, 2021.
	XUE J. Research on key technology of radar emitter signal re-cognition[D]. Nanjing: Nanjing University, 2021.
3	RAKHSHANFAR M. Maximum likelihood approach to classification of digitally frequency-modulated signals[C]//Proc. of the 7th International Symposium on Wireless Communication Systems, 2010: 761-764.
4	吴斌,袁亚博,汪勃.基于记忆因子的连续相位调制信号最大似然调制识别[J].电子与信息学报,2016,38(10):2546-2552.
	WUB,YUANY B,WANGB.Maximum likelihood modulation recognition for continuous phase modulation signals using memory factor[J].Journal of Electronics & Information Technology,2016,38(10):2546-2552.
5	LIC,YANGJ A,LIUH.Modulation recognition algorithm based on cyclic spectrum and ELM-LRF[J].Shipboard Electronic Countermeasure,2020,43(1):52-57, 95.
6	刘腾飞,李贺,孙浩原,等.基于自相关积分循环谱的复合调制雷达信号识别方法[J].中国电子科学研究院学报,2023,18(10):889-895, 929.
	LIUT F,LIH,SUNH Y,et al.Intrapulse composite radar signal recognition method based on autocorrelation integral cyclic spectrum[J].Journal of China Academy of Electronics and Information Technology,2023,18(10):889-895, 929.
7	符颖,王星,周一鹏,等.基于改进半监督朴素贝叶斯的LPI雷达信号识别[J].系统工程与电子技术,2017,39(11):2463-2469. doi: 10.3969/j.issn.1001-506X.2017.11.11
	FUY,WANGX,ZHOUY P,et al.Recognition of radar signals based on revised semi-supervised Naive Bayes[J].Systems Engineering and Electronics,2017,39(11):2463-2469. doi: 10.3969/j.issn.1001-506X.2017.11.11
8	张茜. 基于高阶统计特性的调制信号识别[D]. 成都: 电子科技大学, 2018.
	ZHANG X. Recognition of digital modulation signals based on high-order statistical characteristics[D]. Chengdu: University of Electronic Science and Technology of China, 2018.
9	ZHANGX L,ZHANGJ Z,LUOT Z,et al.Radar signal intrapulse modulation recognition based on a denoising-guided disentangled network[J].Remote Sensing,2022,14(5):1252.
10	HUANGJ,WUB,LIP,et al.Few-shot learning for radar emitter signal recognition based on improved prototypical network[J].Remote Sensing,2022,14(7):1681.
11	HUANGJ,WUB,LIP,et al.Few-shot radar emitter signal recognition based on attention-balanced prototypical network[J].Remote Sensing,2022,14(23):6101.
12	SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2015.
13	KRIZHEASKYA,SUTSKEVERI,HINTONG.ImageNet classification with deep convolutional neural networks[J].Communications of the ACM,2017,60(6):84-90.
14	HU J, SHEN L, SUN G, et al. Squeeze-and-excitation networks[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 7132-7141.
15	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778.
16	WEIS J,QUQ Z,ZENGX F,et al.Self-attention Bi-LSTM networks for radar signal modulation recognition[J].IEEE Trans.on Microwave Theory and Techniques,2021,69(11):5160-5172.
17	LEE D, KIM Y, YOON D. Performance analysis of sequence-based deep learning model for LPI radar waveform recognition in fading channel[C]//Proc. of the 13th International Confe-rence on Information and Communication Technology Convergence, 2022: 2111-2113.
18	徐桂光,王旭东,汪飞,等.基于SE-ResNeXt网络的低信噪比LPI雷达辐射源信号识别[J].系统工程与电子技术,2022,44(12):3676-3684. doi: 10.12305/j.issn.1001-506X.2022.12.11
	XUG G,WANGX D,WANGF,et al.LPI radar emitter signals recognition in low SNR based on SE-ResNeXt network[J].Systems Engineering and Electronics,2022,44(12):3676-3684. doi: 10.12305/j.issn.1001-506X.2022.12.11
19	聂千祁,沙明辉,朱应申.基于改进残差神经网络的雷达信号识别方法[J].系统工程与电子技术,2024,46(10):3356-3364. doi: 10.12305/j.issn.1001-506X.2024.10.13
	NIEQ Q,SHAM H,ZHUY S.Radar signal recognition method based on improved residual neural network[J].Systems Engineering and Electronics,2024,46(10):3356-3364. doi: 10.12305/j.issn.1001-506X.2024.10.13
20	WANGX D,XUG G,YANH,et al.LPI radar signals mo-dulation recognition based on ACDCA-ResNeXt[J].IEEE Access,2023,11,45168-45180.
21	全大英,陈赟,唐泽雨,等.基于双通道卷积神经网络的雷达信号识别[J].上海交通大学学报,2022,56(7):877-885.
	QUAND Y,CHENY,TANGZ Y,et al.Radar signal recognition based on dual channel convolutional neural network[J].Journal of Shanghai Jiaotong University,2022,56(7):877-885.
22	QUAND Y,TANGZ Y,WANGX,et al.LPI radar signal recognition based on dual-channel CNN and feature fusion[J].Symmetry,2022,14(3):570.
23	HUYNHT T,DOANV S,HUAC H,et al.Accurate LPI radar waveform recognition with CWD-TFA for deep convolutional network[J].IEEE Wireless Communications Letters,2021,10(8):1638-1642.
24	HU Z C, HU D X, WANG J T, et al. A pulse accumulation method for LPI radar emitter signal recognition[C]//Proc. of the 4th International Conference on Intelligent Control, Mea-surement and Signal Processing, 2022: 875-880.
25	JIANGW K,LIY,LIAOM M,et al.An improved LPI radar waveform recognition framework with LDC-Unet and SSR-Loss[J].IEEE Signal Processing Letters,2022,29,149-153.
26	HENDERSON A, HARBOUR S, YAKOPCIC C, et al. Spiking Neural networks for LPI radar waveform recognition with neuromorphic computing[C]//Proc. of the IEEE Radar Confe-rence, 2023.
27	PARK D H, BANG J H, PARK J H, et al. A fast and accurate convolutional neural network for LPI radar waveform re-cognition[C]//Proc. of the 19th European Radar Conference, 2022: 89-92.
28	LIUZ P,WANGL D,FENGY T,et al.A recognition method for time-frequency overlapped waveform-agile radar signals based on matrix transformation and multi-scale center point detection[J].Applied Acoustics,2021,175,107855.
29	DAI Y M, GIESEKE F, OEHMCKE S, et al. Attentional feature fusion[C]//Proc. of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021: 3560-3569.
30	LIU Z, LI J G, SHEN Z Q, et al. Learning efficient convolutional networks through network slimming[C]//Proc. of the IEEE International Conference on Computer Vision, 2017: 2755-2763.
31	ZHANGJ M,WANGW,LUC Q,et al.Lightweight deep network for traffic sign classification[J].Annals of Telecommunications,2020,75,369-379.
32	程航,徐树公.基于剪枝与知识蒸馏的轻量级人像抠图方法[J].工业控制计算机,2024,37(1):91-93.
	CHENGH,XUS G.Lightweight human matting method based on pruning and knowledge distillation[J].Industrial Control Computer,2024,37(1):91-93.
33	WANGZ,DUL,LIY.Boosting lightweight CNNs through network pruning and knowledge distillation for SAR target re-cognition[J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,2021,14,8386-8397.
34	ZHANGH Y,YUL,CHENY S,et al.Fast complex-valued CNN for radar jamming signal recognition[J].Remote Sensing,2021,13(15):2867.
35	王海英,张群英,成文海,等.LPI雷达信号调制识别及参数估计研究进展[J].系统工程与电子技术,2024,46(6):1908-1924. doi: 10.12305/j.issn.1001-506X.2024.06.09
	WANGH Y,ZHANGQ Y,CHENGW H,et al.Research progress on LPI radar signal modulation recognition and para-meter estimation[J].Systems Engineering and Electronics,2024,46(6):1908-1924. doi: 10.12305/j.issn.1001-506X.2024.06.09
36	廖正龚,任渊,毕井章.一种频率-相位混合编码的MIMO雷达波形设计[J].现代防御技术,2023,51(6):112-122.
	LIAOZ G,RENY,BIJ Z.A frequency-phase hybrid coded waveform design for MIMO radar[J].Modern Defence Technology,2023,51(6):112-122.
37	付银娟,刘星宇,曾耀平.脉间频率编码脉内复合调频雷达信号设计与分析[J].信号处理,2022,38(10):2053-2063.
	FUY J,LIUX Y,ZENGY P.Design and analysis of interpulse frequency coded and intrapulse composite frequency modulation radar signal[J].Journal of Signal Processing,2022,38(10):2053-2063.
38	胡丹. 低截获概率雷达辐射源信号降噪与检测技术研究[D]. 哈尔滨: 哈尔滨工业大学, 2017.
	HU D. Research on denoising and detection technology of low probability radar radiation source signal[D]. Harbin: Harbin Institute of Technology, 2017.
39	RICHARDS,MARKA.Fundamentals of radar signal processing[M].2nd edNew York:Mcgraw-hill,2005.
40	SHEN L, QUAN D Y, WANG X F, et al. Radar waveform recognition based on time-frequency images under deteriorating channel conditions[C]//Proc. of the IEEE 8th International Conference on Computer and Communications, 2022: 188-192.
41	李欣然,李春国,明钰婷,等.脉冲噪声环境下的调制识别算法[J].信号处理,2024,40(8):1497-1507.
	LIX R,LIC G,MINGY T,et al.Modulation recognition algorithm in impulsive noise environments[J].Journal of Signal Processing,2024,40(8):1497-1507.

序号	网络结构	输出大小
第1部分	2×FF1+MaxPooling	64×27×27
第2部分	3×FF2+MaxPooling	128×13×13
第3部分	3×FF2+MaxPooling	128×6×6
第4部分	3×FF2+MaxPooling	256×3×3
第5部分	3×FF2+AvgPooling	256×1×1
第6部分	最大池化+全连接层	10(分类数量)

雷达波形	参数设置	取值范围
-	信号脉宽/μs	2
	采样率/MHz	500
	起始频率/GHz	1~3
LFM NLFM	信号带宽/MHz	30~200
BPSK/LFM	信号带宽/MHz	30~200
BPSK/LFM	巴克码长度	{7, 11, 13}
MPSK/LFM	信号带宽/MHz	30~200
MPSK/LFM	相位控制数	{4, 5, 6}
StepFrequency/LFM Costas/LFM	子脉冲带宽/MHz	5~30
	跳频间隔/MHz	10~30
	频率编码个数	{4, 5, 6, 7}
StepFrequency/BPSK Costas/BPSK	跳频间隔/MHz	10~30
	频率编码个数	{4, 5, 6, 7}
	巴克码长度	{7, 11, 13}
StepFrequency/MPSK Costas/MPSK	跳频间隔/MHz	10~30
	频率编码个数	{4, 5, 6, 7}
	相位控制数	{4, 5, 6}

SNR/dB	识别准确率	SNR/dB	识别准确率
-8	85.1	-1	99.8
-7	90.9	0	100.0
-6	95.2	1	100.0
-5	98.3	2	100.0
-4	98.9	3	99.8
-3	99.4	4	99.8
-2	99.9	-	-

剪枝率/%	剪枝后每个卷积层通道个数	参数量(×10³)	参数减少率/%	总体识别率/%
0(未剪枝)	[32, 32, 64, 64, 64, 64, 128, 128, 128, 128]	665.9	0	97.36
30	[25, 24, 46, 53, 48, 51, 116, 103, 78, 38]	348.9	47.60	97.47
50	[18, 20, 37, 32, 34, 39, 81, 69, 61, 24,]	176.5	73.49	97.24
70	[14, 16, 21, 19, 22, 24, 43, 44, 36, 10]	63.8	90.42	96.95
80	[6, 11, 16, 15, 13, 13, 26, 27, 31, 8]	27.9	95.81	96.36
82	[6, 8, 16, 14, 13, 12, 25, 24, 24, 7]	22.4	96.64	96.10
84	[6, 7, 16, 10, 13, 10, 18, 24, 22, 7]	17.7	97.34	95.28
86	[5, 7, 13, 10, 12, 9, 12, 20, 22, 6]	13.4	97.99	94.48
88	[5, 5, 11, 10, 10, 8, 8, 18, 21, 3]	9.8	98.53	53.05
90	[5, 4, 11, 9, 9, 6, 5, 10, 21, 3]	6.4	99.04	39.50

α	T	总体识别率/%
0.1	0.5	96.88
0.1	1	96.97
0.1	5	97.14
0.1	10	96.93
0.3	0.5	97.49
0.3	1	97.59
0.3	5	97.68
0.3	10	97.75
0.5	0.5	97.59
0.5	1	97.47
0.5	5	97.64
0.5	10	97.65
0.7	0.5	97.28
0.7	1	96.97
0.7	5	96.95
0.7	10	97.37