Hybrid CNN-Transformer multi-task learning method with dynamic task weighting for radar signal feature recognition

doi:10.11772/j.issn.1001-9081.2026020161

Journal of Computer Applications ›› 2026, Vol. 46 ›› Issue (9): 3017-3024.DOI: 10.11772/j.issn.1001-9081.2026020161

• Frontier and comprehensive applications • Previous Articles

Hybrid CNN-Transformer multi-task learning method with dynamic task weighting for radar signal feature recognition

Hao WU¹, Yulu HUANG²(), Xuan FENG², Qiong ZHANG³, Xiaobo ZHANG³

^1.The 10th Research Institute of China Electronics Technology Group Corporation，Chengdu Sichuan 610036，China
^2.Sichuan Institute of Computer Sciences，Chengdu Sichuan 610041，China
^3.School of Computing and Artificial Intelligence，Southwest Jiaotong University，Chengdu Sichuan 611756，China

Received:2026-02-26 Revised:2026-04-18 Accepted:2026-04-22 Online:2026-05-08 Published:2026-09-10
Contact: Yulu HUANG
About author:WU Hao， born in 1976， Ph. D.， senior engineer. Her research interests include strategic planning management， data mining.
HUANG Yulu， born in 1982， M. S.， senior engineer. Her research interests include multi-task learning， radar signal processing， computer vision.
FENG Xuan， born in 1977， M. S.， professorial senior engineer. His research interests include data fusion， machine learning， intelligent information processing.
ZHANG Qiong， born in 2000， M. S. candidate. Her research interests include time series forecasting， large language models.
ZHANG Xiaobo， born in 1985， Ph. D.， associate professor. His research interests include data mining， machine learning， visualization， computer vision.
Supported by:
Science and Technology Program of Sichuan Province(2025JDRC0028);Key Research and Development Program of Sichuan Province(2024YFFK0410);Regional Innovation Cooperation Program of Sichuan Province(2024YFHZ0066)

雷达信号特征识别的动态任务加权混合CNN-Transformer多任务学习方法

吴昊¹, 黄渝璐²(), 冯暄², 张琼³, 张晓博³

^1.中国电子科技集团公司第十研究所，成都 610036
^2.四川省计算机研究院，成都 610041
^3.西南交通大学计算机与人工智能学院，成都 611756

通讯作者: 黄渝璐
作者简介:吴昊（1976—），女，湖北武汉人，高级工程师，博士，主要研究方向：战略规划管理、数据挖掘
黄渝璐（1982—），女，重庆人，高级工程师，硕士，主要研究方向：多任务学习、雷达信号处理、计算机视觉
冯暄（1977—），男，上海人，正高级工程师，硕士，主要研究方向：数据融合、机器学习、智能信息处理
张琼（2000—），女，河北石家庄人，硕士研究生，主要研究方向：时序预测、大语言模型
张晓博（1985—），男，山西运城人，副教授，博士，CCF高级会员，主要研究方向：数据挖掘、机器学习、可视化、计算机视觉。
基金资助:
四川省科技计划项目(2025JDRC0028);四川省重点研发计划项目(2024YFFK0410);四川省区域创新合作计划项目(2024YFHZ0066)

Abstract

Abstract:

To address the problems in the existing methods for radar signal feature recognition： the poor performance in low Signal-to-Noise Ratio （SNR） environments and the difficulty in handling multiple tasks of classification and parameter estimation simultaneously， a hybrid multi-task learning method fusing Convolutional Neural Network （CNN） and Transformer was proposed. In the method， multi-scale local feature extraction was integrated with global dependency modelling. To further achieve balanced joint optimization， a dynamic task weighting strategy fusing Bayesian uncertainty estimation and task-difficulty awareness was introduced to adjust the loss weights adaptively. Besides， an attention visualization module was developed to enhance model interpretability. Experimental results on the RadChar-Baseline dataset show that compared with baseline models such as IQST （IQ Signal Transformer）， CNN1D （One-dimensional Convolutional Neural Network）， CNN2D （Two-Dimensional Convolutional Neural Network）， ResNet1D （One-Dimensional Residual Network）， and Long Short-Term Memory （LSTM） network， the proposed method achieves a composite score of 0.931； within the SNR range of -20 dB to 20 dB， it has an average composite score of 0.937 with a standard deviation of only 0.108. It can be seen that the proposed method significantly outperforms the comparative methods in both classification and regression tasks， especially under challenging low SNR conditions.

Key words: radar signal feature recognition, multi-task learning, hybrid CNN-Transformer model, dynamic task weighting, attention visualization

摘要：

针对雷达信号特征识别中现有方法在低信噪比（SNR）环境下性能不佳以及难以同时处理分类与参数估计多任务的问题，提出一种融合卷积神经网络（CNN）与Transformer的混合多任务学习方法。该方法将多尺度局部特征提取与全局依赖建模相结合。为进一步实现平衡的联合优化，引入一种融合贝叶斯不确定性估计与任务难度感知的动态任务加权策略，以自适应地调整损失权重。此外，开发注意力可视化模块以增强模型的可解释性。在RadChar-Baseline数据集上的实验结果表明，与IQ信号Transformer （IQST）、一维卷积神经网络（CNN1D）、二维卷积神经网络（CNN2D）、一维残差网络（ResNet1D）及长短期记忆（LSTM）网络等基线模型相比，所提方法的综合得分达到0.931；在SNR为-20 dB~20 dB的范围内，它的平均综合得分达到0.937，性能波动标准差仅为0.108。可见，所提方法在分类与回归任务中均显著优于对比方法，尤其在具有挑战性的低SNR条件下表现突出。

关键词: 雷达信号特征识别, 多任务学习, CNN-Transformer混合模型, 动态任务加权, 注意力可视化

CLC Number:

TP301.6

Hao WU, Yulu HUANG, Xuan FENG, Qiong ZHANG, Xiaobo ZHANG. Hybrid CNN-Transformer multi-task learning method with dynamic task weighting for radar signal feature recognition[J]. Journal of Computer Applications, 2026, 46(9): 3017-3024.

吴昊, 黄渝璐, 冯暄, 张琼, 张晓博. 雷达信号特征识别的动态任务加权混合CNN-Transformer多任务学习方法[J]. 《计算机应用》唯一官方网站, 2026, 46(9): 3017-3024.

Figures/Tables 7

Fig. 1 Structure of proposed method

Tab. 1 Overall performance comparison of different models

模型	分类准确率/%	回归MAE（任务1~4）				综合得分
模型	分类准确率/%	$N p$	$T p w$ /ms	$T P R I$ /ms	$T d$ /ms	综合得分
Hybrid CNN-Transformer	87.48	0.248	0.041	0.039	0.119	0.931
IQST	18.76	4.851	0.784	0.578	0.724	0.155
CNN1D	19.84	4.327	0.122	0.212	0.545	0.288
CNN2D	20.15	4.789	0.408	0.333	0.387	0.237
ResNet1D	81.00	0.403	0.828	0.250	0.084	0.807
LSTM	20.15	4.011	0.011	0.027	0.008	0.376

Tab. 1 Overall performance comparison of different models

模型	分类准确率/%	回归MAE（任务1~4）				综合得分
模型	分类准确率/%	$N p$	$T p w$ /ms	$T P R I$ /ms	$T d$ /ms	综合得分
Hybrid CNN-Transformer	87.48	0.248	0.041	0.039	0.119	0.931
IQST	18.76	4.851	0.784	0.578	0.724	0.155
CNN1D	19.84	4.327	0.122	0.212	0.545	0.288
CNN2D	20.15	4.789	0.408	0.333	0.387	0.237
ResNet1D	81.00	0.403	0.828	0.250	0.084	0.807
LSTM	20.15	4.011	0.011	0.027	0.008	0.376

Tab. 2 Ablation experiment results

模型变体	分类准确率/%	回归MAE	综合得分
Hybrid CNN-Transformer	86.76	0.097 0	0.917 9
w/o Multi-scale CNN	86.46	0.098 9	0.916 2
w/o Transformer	86.68	0.117 5	0.911 5
w/o Dynamic Weights	86.42	0.113 4	0.911 7

Fig. 2 Composite score trends in different SNR levels

Tab. 3 Robustness metrics

模型	综合得分下降幅度	平均得分	标准差
Hybrid CNN-Transformer	0.284	0.937	0.108
LSTM	0.285	0.934	0.110
CNN1D	0.317	0.917	0.124
IQST	0.331	0.847	0.135
CNN2D	0.313	0.909	0.124
ResNet1D	0.296	0.924	0.119

Tab. 4 Computational complexity and inference time comparison

模型	参数量/10⁶	FLOPs/10⁹	推理时间/s	综合得分
Hybrid CNN-Transformer	3.17	1.65	1.28	0.931
LSTM	3.36	0.54	10.22	0.376
CNN1D	0.93	0.55	0.29	0.288
IQST	2.51	0.81	0.37	0.155
CNN2D	0.12	<0.01	0.14	0.237
ResNet1D	4.24	3.49	0.41	0.807

Fig. 3 Average attention heatmaps from 12 attention heads across 3 Transformer encoder layers

References 27

[1]	O’Shea T J， Corgan J， Clancy T C. Deep convolutional networks for radio modulation recognition ［C］// EANN 2016， CCIS 629. Cham： Springer， 2016： 213-226.
[2]	Huynh-The T， Hua C H， Pham Q V， et al. MCNet： efficient convolutional architecture for robust automatic modulation classification ［J］. IEEE Communications Letters， 2020， 24（4）： 811-815.
[3]	Hou K， Du X， Cui G， et al. DCMNet： a supervised learning framework for radar signal modulation recognition ［J］. IEEE Signal Processing Letters， 2025， 32： 2454-2458.
[4]	Lu X， Pan Z， Zhou H. Cross-attention Transformer for coherent detection in radar under low-SNR conditions ［J］. Sensors， 2025， 25（24）： No.7588.
[5]	Jiang M， Quan D， Zhou F， et al. Modulation recognition of radar signals based on multimodal contrastive learning ［J］. IEEE Journal of Microwaves， 2025， 5（5）： 1082-1093.
[6]	Guler B， Geraci G， Jafarkhani H. A multi-task foundation model for wireless channel representation using contrastive and masked autoencoder learning ［PP/OL］. V2. arXiv （2025-10-22）［2025-12-05］. .
[7]	Chen S， Feng Z， Yang S， et al. A generative self-supervised framework for cognitive radio leveraging time-frequency features and attention-based fusion ［J］. IEEE Transactions on Wireless Communications， 2025， 24（3）： 1866-1880.
[8]	Jagannath A， Jagannath J. Multi-task learning approach for automatic modulation and wireless signal classification ［C］// ICC 2021. Piscataway： IEEE， 2021： 1-7.
[9]	Huang Z， Pemasiri A， Denman S， et al. Multi-task learning for radar signal characterisation ［C］// ICASSPW 2023. Piscataway： IEEE， 2023： 1-5.
[10]	Liu L， Wu M， Cheng D， et al. Multi-function working mode recognition based on multi-feature joint learning ［J］. Remote Sensing， 2025， 17（3）： No.521.
[11]	Vaswani A， Shazeer N， Parmar N， et al. Attention is all you need［C］// NeurIPS 2017. Red Hook： Curran Associates Inc.， 2017： 6000-6010.
[12]	Gong Y， Chung Y A， Glass J. AST： audio spectrogram Transformer ［C］// INTERSPEECH 2021. ［S.l.］： International Speech Communication Association， 2021：571-575.
[13]	Wu Y， Meng X， Hu H， et al. Affirm： interactive Mamba with adaptive Fourier filters for long-term time series forecasting ［J］. Proceedings of the AAAI Conference on Artificial Intelligence， 2025， 39（20）： 21599-21607.
[14]	Liu J， Li Y， Zhang Z， et al. Automatic LPI radar modulation recognition using improved VMamba network ［C］// RADAR 2025. Piscataway： IEEE， 2025： 1-6.
[15]	Wu Y， Fioranelli F， Gao C. RadMamba： efficient human activity recognition through radar-based micro-Doppler-oriented Mamba state-space model ［J］. IEEE Transactions on Radar Systems， 2026， 4： 261-272.
[16]	Yang T， Zhang P， Zheng M， et al. WirelessGPT： a generative pre-trained framework for multi-task wireless communication systems ［J］. IEEE Network， 2025， 39（5）： 58-65.
[17]	Aboulfotouh A， Mohammed E， Abou-Zeid H. 6G WavesFM： a foundation model for sensing， communication， and localization［J］. IEEE Open Journal of the Communications Society， 2025， 6：6792-6807.
[18]	Zhou F， Liu C， Zhang H， et al. SpectrumFM： A foundation model for intelligent spectrum management ［J］. IEEE Journal on Selected Areas in Communications， 2025， 44： 4471-4488.
[19]	He H， Qiao H， Chen Y， et al. TimeRadar： a domain-rotatable foundation model for time series anomaly detection ［PP/OL］. arXiv （2026-02-22）［2026-04-17］..
[20]	Kendall A， Gal Y， Cipolla R. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics ［C］// CVPR 2018. Piscataway： IEEE， 2018： 7482-7491.
[21]	Alkhateeb A. DeepMIMO： a generic deep learning dataset for millimeter wave and massive MIMO applications ［PP/OL］. arXiv （2019-02-18）［2026-04-17］. .
[22]	Hoydis J， Cammerer S， Aoudia F A， et al. Sionna： an open-source library for next-generation physical layer research ［PP/OL］. V2. arXiv （2023-03-20）［2026-04-22］..
[23]	张田，杨奎武，魏江宏，等. 面向图像数据的对抗样本检测与防御技术综述［J］. 计算机研究与发展， 2022， 59（6）： 1315-1328.
	Zhang Tian， Yang Kuiwu， Wei Jianghong， et al. Survey on detecting and defending adversarial examples for image data［J］. Journal of Computer Research and Development， 2022， 59（6）： 1315-1328.
[24]	郑海斌，陈晋音，章燕，等. 面向自然语言处理的对抗攻防与鲁棒性分析综述［J］. 计算机研究与发展， 2021， 58（8）： 1727-1750.
	Zheng Haibin， Chen Jinyin， Zhang Yan， et al. Survey of adversarial attack， defense and robustness analysis for natural language processing ［J］. Journal of Computer Research and Development， 2021， 58（8）： 1727-1750.
[25]	徐冬月，田蕴哲，陈康，等. 面向信号调制识别的对抗攻击与防御综述［J］. 计算机研究与发展， 2025， 62（7）： 1713-1737.
	Xu Dongyue， Tian Yunzhe， Chen Kang， et al. Survey on adversarial attack and defense for signal modulation recognition ［J］. Journal of Computer Research and Development， 2025， 62（7）： 1713-1737.
[26]	He K， Zhang X， Ren S， et al. Deep residual learning for image recognition［C］// CVPR 2016. Piscataway： IEEE， 2016： 770-778.
[27]	Zheng Y， Wang J， Huang J. Open space radar specific emitter identification using MSAK-CNN-LSTM network［J］. IET Radar， Sonar & Navigation， 2024， 18（7）： 1080-1093.

Hybrid CNN-Transformer multi-task learning method with dynamic task weighting for radar signal feature recognition

雷达信号特征识别的动态任务加权混合CNN-Transformer多任务学习方法

RichHTML

PDF

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 7

References 27

Related Articles 15

Recommended Articles

Metrics

[1]	Songsen YU, Huang HE, Guopeng XUE, Hengtuo CUI. Quantitation and grading method for ceramic tile chromatic aberration based on improved fractal encoding network [J]. Journal of Computer Applications, 2026, 46(3): 959-968.
[2]	Yi WANG, Yinglong MA. Multi-task social item recommendation method based on dynamic adaptive generation of item graph [J]. Journal of Computer Applications, 2025, 45(8): 2592-2599.
[3]	Longbo YAN, Wentao MAO, Zhihong ZHONG, Lilin FAN. Robust unsupervised multi-task anomaly detection method for defect diagnosis of urban drainage pipe network [J]. Journal of Computer Applications, 2025, 45(6): 1833-1840.
[4]	Weichao DANG, Chujun SONG, Gaimei GAO, Chunxia LIU. Multi-behavior recommendation based on cascading residual graph convolutional network [J]. Journal of Computer Applications, 2025, 45(4): 1223-1231.
[5]	Xinran XU, Shaobing ZHANG, Miao CHENG, Yang ZHANG, Shang ZENG. Bearings fault diagnosis method based on multi-pathed hierarchical mixture-of-experts model [J]. Journal of Computer Applications, 2025, 45(1): 59-68.
[6]	Wei LI, Ling CHEN, Xiuyuan XU, Min ZHU, Jixiang GUO, Kai ZHOU, Hao NIU, Yuchen ZHANG, Shanye YI, Yi ZHANG, Fengming LUO. Interstitial lung disease segmentation algorithm based on multi-task learning [J]. Journal of Computer Applications, 2024, 44(4): 1285-1293.
[7]	Aiguo SHANG, Xinjuan ZHU. Joint approach of intent detection and slot filling based on multi-task learning [J]. Journal of Computer Applications, 2024, 44(3): 690-695.
[8]	Yudan SONG, Jing WANG, Xuehui WANG, Zhaoyang MA, Youfang LIN. Sleep physiological time series classification method based on adaptive multi-task learning [J]. Journal of Computer Applications, 2024, 44(2): 654-662.
[9]	Cunyi LIAO, Yi ZHENG, Weijin LIU, Huan YU, Shouyin LIU. Decoupling-fusing algorithm for multiple tasks with autonomous driving environment perception [J]. Journal of Computer Applications, 2024, 44(2): 424-431.
[10]	Xiao GUO, Yanping CHEN, Ruixue TANG, Ruizhang HUANG, Yongbin QIN. Multi-task learning model for charge prediction with action words [J]. Journal of Computer Applications, 2024, 44(1): 159-166.
[11]	Jianhui HE, Chunlong HU, Xin SHU. Multi-task age estimation method based on multi-peak label distribution learning [J]. Journal of Computer Applications, 2023, 43(5): 1578-1583.
[12]	Ying CHEN, Jiong YU, Jiaying CHEN, Xusheng DU. Cross-layer data sharing based multi-task model [J]. Journal of Computer Applications, 2022, 42(5): 1447-1454.
[13]	Qiming RUAN, Yi GUO, Nan ZHENG, Yexiang WANG. Customs declaration good classification algorithm based on hierarchical multi-task BERT [J]. Journal of Computer Applications, 2022, 42(1): 71-77.
[14]	WU Guoliang, XU Jining. Chinese emergency event extraction method based on named entity recognition task feedback enhancement [J]. Journal of Computer Applications, 2021, 41(7): 1891-1896.
[15]	ZHANG Sun, YIN Chunyong. Sequential multimodal sentiment analysis model based on multi-task learning [J]. Journal of Computer Applications, 2021, 41(6): 1631-1639.