融合时序条件生成对抗网络的小样本雷达对抗侦察数据增强

黄琳玹; 何明浩; 郁春来; 冯明月; 张福群; 张逸楠

doi:10.11999/JEIT250280

融合时序条件生成对抗网络的小样本雷达对抗侦察数据增强

doi: 10.11999/JEIT250280 cstr: 32379.14.JEIT250280

空军预警学院武汉 430019

详细信息

作者简介:
黄琳玹：男，硕士生，研究方向为侦察数据处理

何明浩：男，教授，研究方向为电子对抗侦察

郁春来：男，教授，研究方向为电子对抗侦察

冯明月：男，副教授，研究方向为电子对抗侦察

张福群：男，博士生，研究方向为雷达对抗侦察识别

张逸楠：女，讲师，研究方向为电子对抗侦察

中图分类号: TN957.51
计量
- 文章访问数: 119
- HTML全文浏览量: 66
- PDF下载量: 22
- 被引次数: 0
出版历程
- 收稿日期: 2025-04-16
- 修回日期: 2025-07-22
- 网络出版日期: 2025-07-30

Data Enhancement for Few-shot Radar Countermeasure Reconnaissance via Temporal-Conditional Generative Adversarial Networks

Air Force Early Warning Academy, Wuhan 430019, China

摘要

摘要: 针对雷达对抗侦察中脉冲描述字(PDW)数据稀缺及生成质量不足的问题，该文提出时序条件生成对抗网络(Time-CondGAN)。该方法通过多模态条件生成框架融合脉冲序列时域特征与辐射源分类信息，采用双向门控循环单元 (GRU)监督器与特征匹配损失实现时序连续性与统计分布的双重约束。网络架构包含3个核心模块：(1)条件编码网络将调制类型嵌入为128维特征向量，通过时序维度扩展与潜在噪声拼接，实现类别可控生成；(2)多任务判别器联合执行对抗判别与信号分类，通过共享双向GRU-注意力特征提取层捕捉长程依赖关系；(3)时序-统计联合优化器整合对抗损失、监督损失与特征匹配损失，在对抗训练阶段同步更新生成器与判别器参数。实验表明，Time-CondGAN生成的PDW数据在关键时序参数的KL散度较Time-CondGAN平均降低28.25%，显著提升物理合理性；可视化结果证明模型性能满足要求，消融实验证明改进模块的有效性。在下游任务验证中，生成数据使VGG16与LSTM分类器的识别准确率最高提升37.2%，较Time-CondGAN生成数据平均高8.25%(VGG16)与4.2%(LSTM)。
- 生成对抗网络 /
- 小样本学习 /
- 雷达对抗侦察 /
- 时序数据生成 /
- 模式识别
Abstract: Objective Radar electronic warfare systems rely on Pulse Descriptor Words (PDWs) to represent radar signals, capturing key parameters such as Time of Arrival (TOA), Carrier Frequency (CF), Pulse Width (PW), and Pulse Amplitude (PA). However, in complex electromagnetic environments, the scarcity of PDW data limits the effectiveness of data-driven models in radar pattern recognition. Conventional augmentation methods (e.g., geometric transformations, SMOTE) fall short in addressing three core issues: (1) failure to capture temporal physical laws (e.g., gradual frequency shifts and pulse modulation patterns); (2) distributional inconsistencies (e.g., frequency band overflows and PW discontinuities); and (3) weak coupling between PDWs and modulation types, leading to reduced classification accuracy. This study proposes Time-CondGAN, a temporal-conditional generative adversarial network designed to generate physically consistent and modulation-specific PDW sequences, enhancing few-shot radar reconnaissance performance. Methods Time-CondGAN integrates three core innovations: (1) Multimodal Conditional Generation Framework: A label encoder maps discrete radar modulation types (e.g., VS, HRWS, TWS in Table 2) into 128-dimensional feature vectors. These vectors are temporally expanded and concatenated with latent noise to enable category-controllable generation. The generator architecture (Fig. 4) employs bidirectional Gated Recurrent Units (GRUs) to capture long-range temporal dependencies. (2) Multi-task Discriminator Design: The discriminator (Fig. 5) is designed for joint adversarial discrimination and signal classification. A shared bidirectional GRU with an attention mechanism extracts temporal features, while classification loss constrains the generator to maintain category-specific distributions. (3) Temporal-Statistical Joint Optimization: The training process incorporates multiple loss components: Supervisor Loss (Eq. 7): A bidirectional GRU-based supervisor enforces temporal consistency. Feature Matching Loss (Eq. 9): Encourages alignment between high-level features of real and synthetic signals. Adversarial and Classification Losses (Eqs. (8)–(9)): Promote distribution realism and accurate category separation. A two-stage training strategy is adopted to improve stability, consisting of pre-training (Fig. 2) followed by adversarial training (Fig. 3). Results and Discussions Time-CondGAN demonstrates strong performance across three core dimensions. (1) Physical plausibility is achieved through accurate modeling of radar signal dynamics. Compared with TimeGAN, Time-CondGAN reduces the Kullback-Leibler (KL) divergence by 28.25% on average. Specifically, the PW distribution error decreases by 50.20% (KL = 4.68), the TOA interval error decreases by 20.5% (KL = 0.636), and the CF deviation is reduced by 13.29% (KL = 7.93), confirming the suppression of non-physical signal discontinuities. (2) Downstream task enhancement highlights the model’s few-shot generation capability. With only 10 real samples, classification accuracy improves markedly: VGG16 accuracy increases by 37.2% (from 43.0% to 59.0%) and LSTM accuracy by 28.6% (from 49.0% to 63.0%), both substantially outperforming conventional data augmentation methods. (3) Ablation studies validate the contribution of key modules. Removing the conditional encoder increases the PW KL divergence by 107.4%. Excluding the supervisor loss degrades CF continuity by 76.2%, and omitting feature matching results in a 44.6% misalignment in amplitude distribution. Conclusions This study establishes Time-CondGAN as an effective solution for radar PDW generation under few-shot conditions, addressing three key limitations: temporal fragmentation is resolved through bidirectional GRU supervision, mode collapse is alleviated via multimodal conditioning, and modulation specificity is maintained by classification constraints. The proposed framework offers operational value, enabling over 35% improvement in recognition accuracy under low-intercept conditions. Future work will incorporate radar propagation physics to refine amplitude modeling (current KL = 8.51), adopt meta-learning approaches for real-time adaptation in dynamic battlefield environments, and expand the model to multi-radar cooperative scenarios to support heterogeneous electromagnetic contexts.
- Generative Adversarial Networks (GAN) /
- Few-shot learning /
- Radar countermeasure reconnaissance /
- Temporal data generation /
- Pattern recognition

HTML全文

图 1 时序条件对抗生成网络

下载: 全尺寸图片幻灯片

图 2 预训练流程

下载: 全尺寸图片幻灯片

图 3 对抗训练流程

下载: 全尺寸图片幻灯片

图 4 生成器网络

下载: 全尺寸图片幻灯片

图 5 判别器网络

下载: 全尺寸图片幻灯片

图 6 嵌入器、恢复器与监督器网络

下载: 全尺寸图片幻灯片

图 7 真实数据在VGG16以及LSTM识别性能对比

下载: 全尺寸图片幻灯片

图 8 TimeGAN以及Time-CondGAN在LSTM识别性能

下载: 全尺寸图片幻灯片

图 9 TimeGAN以及Time-CondGAN在VGG16识别性能

下载: 全尺寸图片幻灯片

图 10 Time-CondGAN生成结果

下载: 全尺寸图片幻灯片

图 11 TimeGAN生成结果

下载: 全尺寸图片幻灯片

表 1 实验运行环境

硬件环境		软件环境
名称	型号	名称	版本
CPU处理器	Inter(R) Xenon(R)Silver 4210R	Pycharm	2020.2.3×64
GPU显卡	NVIDIA GeForce RTX 3090	Python	3.11
运行内存	32 GB	Pytorch	1.8.0
操作系统	Win7

下载: 导出CSV

表 2 AESA雷达工作模式参数范围

类别	名称	属性	VS	HRWS	MRWS	TWS	STT	TAS
波形参数	PRI	范围(μs)	4～10	4～10	50～200	5～200	5～200	5～200
波形参数	PRI	类型	固定	固定	组变/抖动	组变/抖动/参差	抖动/参差/滑变/正弦	组变/抖动/参差/滑变/正弦
脉冲参数	PW	范围(μs)	1～3	1～3	1～20	0.1～20	0.1～20	0.1～20
	CF	范围(GHz)	9.5, 10.5	9.5, 10.5	9.5, 10.5	9.5, 10.5	9.5, 10.5	9.5, 10.5
	CF	类型	固定	固定	脉间捷变	脉组捷变	脉组捷变	脉组捷变

下载: 导出CSV

表 3 KL散度对比

参数	Time-CondGAN	TimeGAN	相对改进率(%)
TOA	0.636 0	0.800 0	20.50
幅度	8.510 7	7.810 0	–8.97
载频	7.930 0	9.144 8	13.29
脉宽	4.680 0	9.397 2	50.20

下载: 导出CSV

表 4 识别性能分析表(%)

样本数量	VGG16 (真实数据)	LSTM (真实数据)	VGG16+Time-CondGAN	LSTM +Time-CondGAN	VGG16+TimeGAN	LSTM +TimeGAN
10	42.65	49.17	59.00	63.00	45.10	54.43
20	52.83	53.74	68.00	65.00	58.92	57.92
30	57.80	57.06	74.00	67.00	64.73	62.80
40	61.07	60.90	79.58	70.35	68.17	65.30
50	63.97	64.12	79.31	70.70	70.52	65.30
60	66.62	61.40	85.23	71.58	74.18	66.82
70	67.00	67.82	86.37	74.82	78.08	69.00
80	67.58	65.62	87.10	74.38	78.25	70.95
90	70.00	69.00	87.07	76.45	80.00	73.68

下载: 导出CSV

表 5 消融实验KL散度对比

实验组	TOA	幅度	载频	脉宽
A	0.349 0	7.326 1	11.060 8	9.708 4
B	1.599 8	8.419 1	7.510 8	9.159 6
C	0.172 0	7.431 7	13.981 9	7.471 6
D	0.892 1	12.304 2	9.770 5	7.508 3
完整模型	0.636 0	8.510 7	7.930 0	4.680 0

下载: 导出CSV

参考文献(17)

[1]	陈韬伟, 王会源, 马一鸣, 等. 基于复杂网络建模的雷达辐射源信号脉间特征提取[J]. 火力与指挥控制, 2024, 49(12): 77–84,92. doi: 10.3969/j.issn.1002-0640.2024.12.009. CHEN Taowei, WANG Huiyuan, MA Yiming, et al. Inter-pulse feature extraction of radar emitter signals and its analysis based on complex network modeling[J]. Fire Control & Command Control, 2024, 49(12): 77–84,92. doi: 10.3969/j.issn.1002-0640.2024.12.009.
[2]	陈涛, 邱宝传, 肖易寒, 等. 基于点云分割网络的雷达信号分选方法[J]. 电子与信息学报, 2024, 46(4): 1391–1398. doi: 10.11999/JEIT230622. CHEN Tao, QIU Baochuan, XIAO Yihan, et al. The radar signal deinterleaving method base on point cloud segmentation network[J]. Journal of Electronics & Information Technology, 2024, 46(4): 1391–1398. doi: 10.11999/JEIT230622.
[3]	SHORTEN C and KHOSHGOFTAAR T M. A survey on image data augmentation for deep learning[J]. Journal of Big Data, 2019, 6(1): 60. doi: 10.1186/s40537-019-0197-0.
[4]	CHAWLA N V, BOWYER K W, HALL L O, et al. SMOTE: Synthetic minority over-sampling technique[J]. Journal of Artificial Intelligence Research, 2002, 16: 321–357. doi: 10.1613/jair.953.
[5]	GOODFELLOW I J, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[C]. The 28th International Conference on Neural Information Processing Systems, Montreal, Canada, 2014: 2672–2680.
[6]	MIRZA M and OSINDERO S. Conditional generative adversarial nets[Z]. arXiv preprint arXiv: 1411.1784, 2014. doi: 10.48550/arXiv.1411.1784.
[7]	YOON J, JARRETT D, and VAN DER SCHAAR M. Time-series generative adversarial networks[C]. The 33rd International Conference on Neural Information Processing Systems, Vancouver, Canada, 2019: 494.
[8]	KHAN S and KUMAR V. A novel hybrid GRU-CNN and residual bias (RB) based RB-GRU-CNN models for prediction of PTB Diagnostic ECG time series data[J]. Biomedical Signal Processing and Control, 2024, 94: 106262. doi: 10.1016/j.bspc.2024.106262.
[9]	WANG Zhongqi, SONG Chong, JIAO Zekun, et al. Synthetic aperture radar deep statistical imaging through diffusion generative model conditional inference[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5227717. doi: 10.1109/TGRS.2024.3498442.
[10]	ANG Yihao, HUANG Qiang, BAO Yifan, et al. TSGBench: Time series generation benchmark[J]. Proceedings of the VLDB Endowment, 2023, 17(3): 305–318. doi: 10.14778/3632093.3632097.
[11]	ESKANDARINASAB M R, HAMDI S M, and BOUBRAHIMI S F. ChronoGAN: Supervised and embedded generative adversarial networks for time series generation[C]. 2024 International Conference on Machine Learning and Applications (ICMLA), Miami, USA, 2024: 567–574. doi: 10.1109/ICMLA61862.2024.00083.
[12]	TANG Haocheng, LONG Jing, and WANG Junmei. Auxiliary discrminator sequence generative adversarial networks (ADSeqGAN) for few sample molecule generation[Z]. arXiv preprint arXiv: 2502.16446, 2025. doi: 10.48550/arXiv.2502.16446.
[13]	CHOWDHURY M A, MODEKWE G, and LU Qiugang. Lithium-ion battery capacity prediction via conditional recurrent generative adversarial network-based time-series regeneration[Z]. arXiv preprint arXiv: 2503.12258, 2025. doi: 10.48550/arXiv.2503.12258.
[14]	RUAN Shulan, ZHANG Yong, ZHANG Kun, et al. DAE-GAN: Dynamic aspect-aware Gan for text-to-image synthesis[C]. The IEEE/CVF International Conference on Computer Vision, Montreal, Canada, 2021: 13940–13949. doi: 10.1109/ICCV48922.2021.01370.
[15]	YANG Yue, ZHANG Zhuo, MAO Wei, et al. Radar target recognition based on few-shot learning[J]. Multimedia Systems, 2023, 29(5): 2865–2875. doi: 10.1007/s00530-021-00832-3.
[16]	史松昊, 王晓丹, 杨春晓, 等. 基于跨域小样本学习的SAR图像目标识别方法[J]. 计算机科学, 2024, 51(6A): 230800136. doi: 10.11896/jsjkx.230800136. SHI Songhao, WANG Xiaodan, YANG Chunxiao, et al. SAR image target recognition based on cross domain few shot learning[J]. Computer Science, 2024, 51(6A): 230800136. doi: 10.11896/jsjkx.230800136.
[17]	王玉冰, 程嗣怡, 周一鹏, 等. 基于DS证据理论的机载火控雷达空空工作模式判定[J]. 现代雷达, 2017, 39(5): 79–84. doi: 10.16592/j.cnki.1004-7859.2017.05.017. WANG Yubing, CHENG Siyi, ZHOU Yipeng, et al. Air-to-air operation modes recognition of airborne fire control radar based on DS evidence theory[J]. Modern Radar, 2017, 39(5): 79–84. doi: 10.16592/j.cnki.1004-7859.2017.05.017.