零记忆增量学习的复合有源干扰识别

吴振华; 崔金鑫; 曹宜策; 张强; 张磊; 杨利霞

doi:10.11999/JEIT240521

零记忆增量学习的复合有源干扰识别

doi: 10.11999/JEIT240521

吴振华^{1, 2},
崔金鑫¹,
曹宜策^1, ,,
张强³,
张磊⁴,
杨利霞¹

1.
安徽大学信息材料与智能感知安徽省实验室合肥 230601
2.
中国电子科技集团公司第三十八研究所合肥 230088
3.
天基综合信息系统全国重点实验室北京 100094
4.
中山大学电子与通信工程学院深圳 518107

基金项目: 国家自然科学基金(62201007, 62401007)，中国博士后科学基金(2020M681992)，安徽省自然科学基金(2308085QF199)

详细信息

作者简介:
吴振华：男，副教授，研究方向为雷达成像、雷达信号处理、雷达智能干扰对抗

崔金鑫：男，硕士生，研究方向为雷达干扰识别、雷达信号处理

曹宜策：女，讲师，研究方向为遥感图像的智能解译、雷达信号处理、计算机视觉

张强：男，研究员，研究方向为天基信号处理

张磊：男，教授，研究方向为雷达信号处理

杨利霞：男，教授，研究方向为电磁散射与逆散射、电波传播及天线理论与设计、计算电磁学

通讯作者:
曹宜策　yccao@ahu.edu.cn

中图分类号: TN974
计量
- 文章访问数: 203
- HTML全文浏览量: 94
- PDF下载量: 36
- 被引次数: 0
出版历程
- 收稿日期: 2024-06-25
- 修回日期: 2024-11-07
- 网络出版日期: 2024-11-13
- 刊出日期: 2025-01-31

Compound Active Jamming Recognition for Zero-memory Incremental Learning

WU Zhenhua^{1, 2},
CUI Jinxin¹,
CAO Yice^{1
, ,},
ZHANG Qiang³,
ZHANG Lei⁴,
YANG Lixia¹

1.
Information Materials and Intelligent Sensing Laboratory of Anhui Province, Anhui University, Hefei 230601, China
2.
Thirty-eighth Research Institute of China Electronics Technology Group Corporation, Hefei 230088, China
3.
National Key Laboratory of Space Integrated Information System, Beijing 100094, China
4.
School of Electronics and Communication Engineering, Sun Yat-sen University, Shenzhen 518107, China

Funds: The National Natural Science Foundation of China (62201007, 62401007), China Postdoctoral Science Foundation (2020M681992), The Natural Science Foundation of Anhui (2308085QF199)

摘要

摘要: 非完备、高动态有源干扰对抗作战环境下，现阶段针对库内多类型单一有源干扰样本所优化训练的静态模型，在面对库外类型多样、参数多变、组合方式多元的复合干扰时，模型无法快速更新且难以应对测试样本数非均衡问题。针对此问题，该文提出一种基于零记忆增量学习的雷达复合有源干扰识别方法。首先，利用元学习训练模式对库内单一干扰进行原型学习，训练出高效的特征提取器，使其具备对库外复合干扰特征有效提取能力。进而，基于超维空间和余弦相似度计算，构建零记忆增量学习网络(ZMILN)，将复合干扰原型向量映射到超维空间并存储，从而实现识别模型动态更新。此外，为解决样本数非均衡下复合干扰识别问题，设计直推式信息最大化(TIM)测试模块，通过在互信息损失函数中加入散度约束，对识别模型进一步强化训练以应对非均衡测试样本。实验结果表明，该文所提方法在非均衡测试条件下对4种单一干扰和7种复合干扰进行增量学习后，平均识别准确率达到了93.62%。该方法通过对库内多类型单一干扰知识充分提取，实现对多种组合条件下库外复合干扰的快速动态识别。
- 雷达有源干扰 /
- 零记忆增量学习 /
- 非均衡 /
- 直推式信息最大化 /
- 复合干扰识别
Abstract: Objective: In contemporary warfare, radar systems serve a crucial role as vital instruments for detection and tracking. Their performance is essential, often directly impacting the progression and outcome of military engagements. As these systems operate in complex and hostile environments, their susceptibility to adversarial interference becomes a significant concern. Recent advancements in active jamming techniques, particularly compound active jamming, present considerable threats to radar systems. These jamming methods are remarkably adaptable, employing a range of signal types, parameter variations, and combination techniques that complicate countermeasures. Not only do these jamming signals severely impair the radar’s ability to detect and track targets, but they also exhibit rapid adaptability in high-dynamic combat scenarios. This swift evolution of jamming techniques renders traditional radar jamming recognition models ineffective, as they struggle to address the fast-changing nature of these threats. To counter these challenges, this paper proposes a novel incremental learning method designed for recognizing compound active jamming in radar systems. This innovative approach seeks to bridge the gaps of existing methods when confronted with incomplete and dynamic jamming conditions typical of adversarial combat situations. Specifically, it tackles the challenge of swiftly updating models to identify novel out-of-database compound jamming while mitigating the performance degradation caused by imbalanced sample distributions. The primary objective is to enhance the adaptability and reliability of radar systems within complex electronic warfare environments, ensuring robust performance against increasingly sophisticated and unpredictable jamming techniques. Methods: The proposed method commences with prototypical learning within a meta-learning framework to achieve efficient feature extraction. Initially, a feature extractor is trained utilizing in-database single jamming signals. This extractor is thoroughly designed to proficiently capture the features of out-of-database compound jamming signals. Subsequently, a Zero-Memory Incremental Learning Network (ZMILN) is developed, which incorporates hyperdimensional space and cosine similarity techniques. This network facilitates the mapping and storage of prototype vectors for compound jamming signals, thereby enabling the dynamic updating of the recognition model. To address the challenges associated with imbalanced test sample distributions, a Transductive Information Maximization (TIM) testing module is introduced. This module integrates divergence constraints into the mutual information loss function, refining the recognition model to optimize its performance across imbalanced datasets. The implementation begins with a comprehensive modeling of radar active jamming signals. Linear Frequency Modulation (LFM) signals, frequently utilized in contemporary radar systems, are chosen as the foundation for the transmitted radar signals. The received signals are modeled as a blend of target echo signals, jamming signals, and noise. Various categories of radar active jamming, including suppression jamming and deceptive jamming, are classified, and their composite forms are examined. For feature extraction, a five-layer Convolutional Neural Network (CNN) is employed. This CNN is specifically designed to transform input radar jamming time-frequency image samples into a hyperdimensional feature space, generating 512-dimensional prototype vectors. These vectors are then stored within the prototype space, with each jamming category corresponding to a distinct prototype vector. To enhance classification accuracy and efficiency, a quasi-orthogonal optimization strategy is utilized to improve the spatial arrangement of these prototype vectors, thereby minimizing overlap and confusion between different categories and increasing the precision of jamming signal recognition. The ZMILN framework addresses two primary challenges in recognizing compound jamming signals: the scarcity of new-category samples and the limitations inherent in existing models when it comes to identifying novel categories. By integrating prototypical learning with hyperdimensional space techniques, the ZMILN enables generalized recognition from in-database single jamming signals to out-of-database compound jamming. To further enhance model performance in the face of imbalanced sample conditions, the TIM module maximizes information gain by partitioning the test set into supervised support and unsupervised query sets. The ZMILN model is subsequently fine-tuned using the support set, followed by unsupervised testing on the query set. During the testing phase, the model computes the cosine similarity between the test samples and the prototype vectors, ultimately yielding the final recognition results. Results and Discussions: The proposed method exhibits notable effectiveness in the recognition of radar compound active jamming signals. Experimental results indicate an average recognition accuracy of 93.62% across four single jamming signals and seven compound jamming signals under imbalanced test conditions. This performance significantly exceeds various baseline incremental learning methods, highlighting the superior capabilities of the proposed approach in the radar jamming recognition task. Additionally, t-distributed Stochastic Neighbor Embedding (t-SNE) visualization experiments present the distribution of jamming features at different stages of incremental learning, further confirming the method’s effectiveness and robustness. The experiments simulate a realistic radar jamming recognition scenario by categorizing “in-database” jamming as single types included in the base training set, and “out-of-database” jamming as novel compound types that emerge during the incremental training phase. This configuration closely resembles real-world operational conditions, where radar systems routinely encounter new and evolving jamming techniques. Quantitative performance metrics, including accuracy and performance degradation rates, are utilized to assess the model’s capacity to retain knowledge of previously learned categories while adapting to new jamming types. Accuracy is computed at each incremental learning stage to evaluate the model’s performance on both old and new categories. Furthermore, the performance degradation rate is calculated to measure the extent of knowledge retention, with lower degradation rates indicative of stronger retention of prior knowledge throughout the learning process. Conclusions: In conclusion, the proposed Zero-Memory Incremental Learning method for recognizing radar compound active jamming is highly effective in addressing the challenges posed by rapidly evolving and complex radar jamming techniques. By leveraging a comprehensive understanding of individual jamming signals, this method facilitates swift and dynamic recognition of out-of-database compound jamming across diverse and high-dynamic conditions. This approach not only enhances the radar system’s capabilities in recognizing novel compound jamming but also effectively mitigates performance degradation resulting from imbalanced sample distributions. Such advancements are essential for improving the adaptability and reliability of radar systems in complex electronic warfare environments, where the nature of jamming signals is in constant flux. Additionally, the proposed method holds significant implications for other fields facing incremental learning challenges, particularly those involving imbalanced data and rapidly emerging categories. Future research will focus on exploring open-set recognition models, further enhancing the cognitive recognition capabilities of radar systems in fully open and highly dynamic adversarial environments. This work lays the groundwork for developing more agile cognitive closed-loop recognition systems, ultimately contributing to more resilient and adaptable radar systems capable of effectively managing complex electronic warfare scenarios.
- Radar active jamming /
- Zero-memory incremental learning /
- Imbalance /
- Transductive Information Maximization(TIM) /
- Compound jamming recognition

HTML全文

图 1 雷达回波及干扰信号的时频谱图

下载: 全尺寸图片幻灯片

图 2 零记忆增量学习网络

下载: 全尺寸图片幻灯片

图 3 原型学习网络详细结构图

下载: 全尺寸图片幻灯片

图 4 原型空间更新示意图

下载: 全尺寸图片幻灯片

图 5 不同方法下的干扰识别准确率曲线

下载: 全尺寸图片幻灯片

图 6 1-way 5-shot设置下的t-SNE可视化图

下载: 全尺寸图片幻灯片

图 7 不同基础设置下的干扰识别性能曲线

下载: 全尺寸图片幻灯片

表 1 雷达干扰参数

信号	参数	数值范围
LFM	信号宽度带宽采样频率	10 μs 50 μs 125 MHz
SMSP	干扰个数	3～7
SNJ	转发个数切片长度占空比	3～5 10 μs 0.5～0.8
DFTJ	假目标个数假目标时延	3～7 1～10 μs
MISRJ	转发个数切片长度占空比	3～5 10 μs 0.5～0.8

下载: 导出CSV

表 2 增量干扰数据集配置

阶段	序号	干扰名称	训练样本个数	单次测试样本个数
基础	1	SMSP	100	1
	2	SNJ	100	2
	3	DFTJ	100	2
	4	MISRJ	100	8
增量	5	SNJ+SMSP	5	12
	6	SMSP+MISRJ	5	5
	7	DFTJ+MISRJ	5	9
	8	SMSP+DFTJ	5	5
	9	SNJ+DFTJ	5	12
	10	SNJ+MISRJ	5	7
	11	SNJ+SMSP+DFTJ	5	14

下载: 导出CSV

表 3 TIM模块对模型的影响(%)

测试样本分布	TIM	基础阶段准确率	增量阶段准确率							平均准确率	性能下降率
测试样本分布	TIM	1	2	3	4	5	6	7	8	平均准确率	性能下降率
均衡	√	100.00	100.00	97.95	97.60	96.61	95.72	94.88	94.62	97.17	5.38
均衡	$ \times $	100.00	97.42	92.94	91.70	90.42	89.21	87.61	85.72	91.87	14.28
非均衡	√	100.00	100.00	98.95	97.60	96.61	94.72	93.88	93.62	96.92	6.38
非均衡	$ \times $	99.85	95.22	93.71	87.83	86.44	84.72	72.51	80.66	87.61	19.19

下载: 导出CSV

表 4 在1-way 5-shot设置下不同方法的干扰识别结果

方法	基础阶段准确率(%)	增量阶段准确率(%)							平均准确率 (%)	性能下降率 (%)	训练平均时间 (min)	测试平均时间 (s)
方法	1	2	3	4	5	6	7	8	平均准确率 (%)	性能下降率 (%)	训练平均时间 (min)	测试平均时间 (s)
Ft-CNN^[8]	100.00	92.22	83.33	74.16	67.77	61.33	55.55	51.55	73.23	48.45	50.59	15.73
iCaRL^[24]	100.00	91.14	85.72	85.83	83.51	82.66	80.75	78.47	86.01	21.53	185.15	20.79
TOPIC^[16]	99.31	89.77	83.54	84.91	84.67	81.03	78.75	75.98	84.74	23.33	141.96	19.86
FACT^[26]	97.55	97.77	93.80	92.57	91.29	88.74	85.45	83.12	91.28	14.43	98.15	12.02
CEC^[23]	98.44	96.74	93.94	92.70	90.61	88.72	85.88	84.62	91.45	13.82	169.64	16.95
F2M^[25]	100.00	95.71	93.44	87.70	86.41	84.72	82.45	80.74	88.89	19.26	78.96	9.93
本文(ZMILN)	100.00	100.00	98.95	97.60	96.61	94.72	93.88	93.62	96.92	6.38	62.45	36.95

下载: 导出CSV

参考文献(26)

[1]	魏毅寅, 杨文华. 海战场典型干扰对抗场景及反舰导弹应对策略研究[J]. 战术导弹技术, 2020(5): 1–8. doi: 10.16358/j.issn.1009-1300.2020.1.538. WEI Yiyin and YANG Wenhua. Study on typical jamming scenes in naval battle field and countermeasures of anti-ship missile[J]. Tactical Missile Technology, 2020(5): 1–8. doi: 10.16358/j.issn.1009-1300.2020.1.538.
[2]	ZHANG Xiang, LAN Lan, ZHU Shengqi, et al. Intelligent suppression of interferences based on reinforcement learning[J]. IEEE Transactions on Aerospace and Electronic Systems, 2024, 60(2): 1400–1415. doi: 10.1109/TAES.2023.3336643.
[3]	崔国龙, 余显祥, 魏文强, 等. 认知智能雷达抗干扰技术综述与展望[J]. 雷达学报, 2022, 11(6): 974–1002. doi: 10.12000/JR22191. CUI Guolong, YU Xianxiang, WEI Wenqiang, et al. An overview of antijamming methods and future works on cognitive intelligent radar[J]. Journal of Radars, 2022, 11(6): 974–1002. doi: 10.12000/JR22191.
[4]	周红平, 王子伟, 郭忠义. 雷达有源干扰识别算法综述[J]. 数据采集与处理, 2022, 37(1): 1–20. doi: 10.16337/j.1004-9037.2022.01.001. ZHOU Hongping, WANG Ziwei, and GUO Zhongyi. Overview on recognition algorithms of radar active jamming[J]. Journal of Data Acquisition and Processing, 2022, 37(1): 1–20. doi: 10.16337/j.1004-9037.2022.01.001.
[5]	QU Qizhe, WEI Shunjun, LIU Shan, et al. JRNet: Jamming recognition networks for radar compound suppression jamming signals[J]. IEEE Transactions on Vehicular Technology, 2020, 69(12): 15035–15045. doi: 10.1109/TVT.2020.3032197.
[6]	ZHANG Jiaxiang, LIANG Zhennan, ZHOU Chao, et al. Radar compound jamming cognition based on a deep object detection network[J]. IEEE Transactions on Aerospace and Electronic Systems, 2023, 59(3): 3251–3263. doi: 10.1109/TAES.2022.3224695.
[7]	LV Qinzhe, QUAN Yinghui, FENG Wei, et al. Radar deception jamming recognition based on weighted ensemble CNN with transfer learning[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5107511. doi: 10.1109/TGRS.2021.3129645.
[8]	DU Jinbiao, FAN Weiwei, GONG Chen, et al. Aggregated-attention deformable convolutional network for few-shot SAR jamming recognition[J]. Pattern Recognition, 2024, 146: 109990. doi: 10.1016/j.patcog.2023.109990.
[9]	LUO Zhenyu, CAO Yunhe, YEO T S, et al. Few-shot radar jamming recognition network via time-frequency self-attention and global knowledge distillation[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5105612. doi: 10.1109/TGRS.2023.3280322.
[10]	ZHOU Hongping, WANG Lei, GUO Zhongyi. Recognition of radar compound jamming based on convolutional neural network[J]. IEEE Transactions on Aerospace and Electronic Systems, 2023, 59(6): 7380–7394. doi: 10.1109/TAES.2023.3288080.
[11]	LI Boran, ZHANG Lei, DAI Jingwei, et al. FETTrans: Analysis of compound interference identification based on bidirectional dynamic feature adaptation of improved transformer[J]. IEEE Access, 2022, 10: 66321–66331. doi: 10.1109/ACCESS.2022.3182010.
[12]	MENG Yunyun, YU Lei, and WEI Yinsheng. Multi-label radar compound jamming signal recognition using complex-valued CNN with jamming class representation fusion[J]. Remote Sensing, 2023, 15(21): 5180. doi: 10.3390/rs15215180.
[13]	ZHOU Hongping, WANG Lei, MA Minghui, et al. Compound radar jamming recognition based on signal source separation[J]. Signal Processing, 2024, 214: 109246. doi: 10.1016/j.sigpro.2023.109246.
[14]	LV Qinzhe, FAN Hanxin, LIU Junliang, et al. Multilabel deep learning-based lightweight radar compound jamming recognition method[J]. IEEE Transactions on Instrumentation and Measurement, 2024, 73: 2521115. doi: 10.1109/TIM.2024.3400337.
[15]	KONG Yukai, XIA Senlin, DONG Luxin, et al. Compound jamming recognition via contrastive learning for distributed MIMO radars[J]. IEEE Transactions on Vehicular Technology, 2024, 73(6): 7892–7907. doi: 10.1109/TVT.2024.3358996.
[16]	TAO Xiaoyu, HONG Xiaopeng, CHANG Xinyuan, et al. Few-shot class-incremental learning[C]. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020: 12183–12192. doi: 10.1109/CVPR42600.2020.01220.
[17]	LI Bin, CUI Zongyong, WANG Haohan, et al. SAR incremental automatic target recognition based on mutual information maximization[J]. IEEE Geoscience and Remote Sensing Letters, 2024, 21: 4005305. doi: 10.1109/LGRS. 2024.3368063.
[18]	SERBES A. On the estimation of LFM signal parameters: Analytical formulation[J]. IEEE Transactions on Aerospace and Electronic Systems, 2018, 54(2): 848–860. doi: 10.1109/TAES.2017.2767978.
[19]	WANG J X. Meta-learning in natural and artificial intelligence[J]. Current Opinion in Behavioral Sciences, 2021, 38: 90–95. doi: 10.1016/j.cobeha.2021.01.002.
[20]	SANTORO A, BARTUNOV S, BOTVINICK M, et al. Meta-learning with memory-augmented neural networks[C]. The 33rd International Conference on International Conference on Machine Learning, New York, USA, 2016: 1842-1850.
[21]	KARUNARATNE G, SCHMUCK M, LE GALLO M, et al. Robust high-dimensional memory-augmented neural networks[J]. Nature Communications, 2021, 12(1): 2468. doi: 10.1038/s41467-021-22364-0.
[22]	VEILLEUX O, BOUDIAF M, PIANTANIDA P, et al. Realistic evaluation of transductive few-shot learning[C]. Proceedings of the 35th International Conference on Neural Information Processing Systems, 2021: 711.
[23]	ZHANG Chi, SONG Nan, LIN Guosheng, et al. Few-shot incremental learning with continually evolved classifiers[C]. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 12455–12464. doi: 10.1109/CVPR46437.2021.01227.
[24]	REBUFFI S A, KOLESNIKOV A, SPERL G, et al. iCaRL: Incremental classifier and representation learning[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 2001–2010. doi: 10.1109/cvpr.2017.587.
[25]	SHI Guangyuan, CHEN Jiaxin, ZHANG Wenlong, et al. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima[C]. The 35th International Conference on Neural Information Processing Systems, 2021: 517.
[26]	ZHOU Dawei, WANG Fuyun, YE Hanjia, et al. Forward compatible few-shot class-incremental learning[C]. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 9046–9056. doi: 10.1109/cvpr52688.2022.00884.