面向无人机射频识别的轻量化双流卷积网络特征融合方法

董鹏宇; 向新; 吕思婷; 梁源; 王瑞; 毛虎

doi:10.11999/JEIT260464

面向无人机射频识别的轻量化双流卷积网络特征融合方法

doi: 10.11999/JEIT260464 cstr: 32379.14.JEIT260464

空军工程大学西安 710038

详细信息

作者简介:
董鹏宇：男，讲师，研究方向为通信侦察与智能信号处理、无人机射频信号检测与识别

向新：男，教授，研究方向为通信侦察与智能信号处理、无线通信技术

吕思婷：女，讲师，研究方向为软件无线电、深度学习技术

梁源：男，副教授，研究方向为通信侦察与智能信号处理、无线通信技术

王瑞：女，讲师，研究方向为软件无线电、深度学习技术

毛虎：男，讲师，研究方向为通信侦察与智能信号处理、无线通信技术

通讯作者:
王瑞　15353724115@163.com

中图分类号: TP391.45
计量
- 文章访问数: 8
- HTML全文浏览量: 1
- PDF下载量: 1
- 被引次数: 0
出版历程
- 收稿日期: 2026-04-30
- 修回日期: 2026-07-31
- 录用日期: 2026-08-10
- 网络出版日期: 2026-08-20

A Lightweight Dual-Stream Convolutional Network Feature Fusion Method for UAV RF Recognition

Air Force Engineering University, Xi’an, 710038, China

摘要

摘要: 针对低信噪比环境下传统射频指纹识别方法特征提取鲁棒性较弱，且现有深度学习模型参数量庞大、无法适配战术边缘设备部署的问题，本文提出一种面向无人机射频识别的轻量化双流卷积网络特征融合方法。该方法构建静态与动态双分支并行架构，依托元素级相加融合策略实现特征互补优化。静态分支以短时傅里叶变换时频图为输入，通过卷积模块提取信号频谱纹理特征；动态分支引入时间梯度一阶差分运算，有效抑制静态背景噪声，提升低信噪比下信号时频特征的表征对比度。基于DroneRF公开数据集的实验结果表明，该方法识别准确率可达95.65%，相较单流网络提升5个百分点，平均信噪比提升1 dB以上；模型参数量仅0.58 M，可满足战术边缘设备实时识别需求。消融实验与特征可视化结果证明，静态频谱特征与动态时间梯度特征的互补特性，是提升无人机射频识别性能的核心关键。
- 无人机射频识别 /
- 轻量化双流卷积网络 /
- 特征融合 /
- 时间梯度 /
- 短时傅里叶变换
Abstract: Objective With the rapid proliferation of Unmanned Aerial Vehicles (UAVs) and the escalating demand for airspace security, radio frequency (RF) fingerprint recognition has emerged as a pivotal technology for identifying non-cooperative UAVs. However, existing methods grapple with significant challenges, including poor robustness in low signal-to-noise ratio (SNR) environments and prohibitive computational complexity, which severely hinder their deployment on resource-constrained tactical edge devices. To address these critical limitations, this paper proposes a novel lightweight dual-stream convolutional network tailored for UAV RF recognition. This network is designed to extract static spectral texture features and dynamic temporal gradient features in parallel, complemented by a meticulously crafted lightweight feature fusion strategy. Methods The proposed network architecture is ingeniously designed to process RF signals. The input signal undergoes a Short-Time Fourier Transform (STFT) to generate a two-dimensional spectrogram, which serves as the primary input. The network is bifurcated into two parallel streams: a static stream and a dynamic stream. The static stream is engineered to capture the inherent static spectral patterns and energy distributions within the STFT spectrogram. It comprises a series of stacked convolutional blocks, each integrating convolutional layers, batch normalization, and ReLU activation functions, followed by max-pooling layers to progressively downsample the feature maps and increase the channel depth. Conversely, the dynamic stream is dedicated to enhancing feature discriminability, particularly in low-SNR scenarios. It begins by computing the temporal gradient of the input spectrogram, effectively suppressing static background noise and accentuating dynamic signal variations. This gradient map is then processed by a symmetric set of convolutional blocks, mirroring the structure of the static stream. To maintain model efficiency, an element-wise addition fusion strategy is employed to integrate the features from both streams, ensuring a balance between feature complementarity and computational overhead. The fused features are subsequently fed into a classification head, consisting of an adaptive average pooling layer, dropout layers for regularization, and fully connected layers to produce the final classification output. Extensive experiments are conducted on the publicly available DroneRF dataset, encompassing ablation studies to dissect the contribution of each component, comparative analyses of various fusion strategies, and rigorous evaluations of the model’s lightweight characteristics. Results and Discussions The experimental results unequivocally demonstrate the efficacy of the proposed method. The dual-stream network achieves a remarkable 95.65% accuracy on the test set, representing a substantial 5 percentage point improvement over the best-performing single-stream network. A critical analysis reveals that the temporal gradient operation contributes significantly to this enhancement by improving the average SNR by 1 dB, thereby bolstering feature discriminability in challenging low-SNR environments. Furthermore, the model’s lightweight design is a standout feature, with a mere 0.58 million parameters, making it eminently suitable for deployment on tactical edge devices. Ablation studies and feature visualization analyses provide compelling evidence for the complementary nature of static and dynamic features. The static stream adeptly captures broad spectral contours, while the dynamic stream focuses on fine-grained temporal variations. The element-wise addition fusion strategy proves superior, outperforming other approaches like feature concatenation and attention-based fusion in terms of both performance and computational efficiency, thereby validating the rationale behind the lightweight design. Conclusions This paper presents a comprehensive solution to the challenges of UAV RF recognition in complex environments by proposing a lightweight dual-stream convolutional network. The method effectively enhances recognition accuracy and robustness through the synergistic combination of dual-stream feature extraction and the SNR-enhancing properties of temporal gradient features, all while maintaining a lightweight architecture suitable for edge deployment. The proposed approach offers a significant advancement in the field, providing a robust and efficient solution for UAV identification. Future research endeavors will focus on further enhancing the model’s adaptability to complex electromagnetic environments, incorporating the effects of sensor noise, and extending the framework to multi-UAV cooperative scenarios.
- UAV RF Recognition /
- Lightweight Dual-Stream Convolutional Network /
- Feature Fusion /
- Temporal Gradient /
- Short-Time Fourier Transform (STFT)

HTML全文

图 1 轻量化双流卷积网络特征融合识别框架

下载: 全尺寸图片幻灯片

图 2 LDSC-Net 特征维度演化图

下载: 全尺寸图片幻灯片

图 3 消融实验得到的总体识别准确率结果

下载: 全尺寸图片幻灯片

图 4 消融实验得到的分类别识别准确率结果

下载: 全尺寸图片幻灯片

图 5 动、静态流对识别效果的特征互补性可视化

下载: 全尺寸图片幻灯片

图 6 不同特征融合策略下识别性能对比

下载: 全尺寸图片幻灯片

图 7 动静态双流特征激活可视化

下载: 全尺寸图片幻灯片

图 8 时间梯度特征的时频表征

下载: 全尺寸图片幻灯片

表 1 仿真实验参数设置

类型	参数名称	参数	类型	参数名称	参数	类型	参数名称	参数
样本	STFT窗函数	Hamming	模型	输入尺寸	1×128×128	训练	Batch Size	32
	FFT点数	1024		卷积核	3×3		训练轮数	100
	帧移	512		池化	MaxPool2d		优化器	Adam
	STFT输出尺寸	128×128		激活函数	ReLU		初始学习率	0.0002

下载: 导出CSV

表 2 不同模型性能比较

Model	Parameters	Accuracy	F1-score	ITPA (ms)
ResNet-18	11.17M	95.11%	0.96	5.49
EfficientNetV1	7.03M	95.65%	0.96	18.60
MobileNetV2	2.23M	93.48%	0.93	11.79
ShuffleNetV2	1.26M	89.86%	0.90	14.50
SqueezeNet	0.73M	94.75%	0.95	8.01
LDSC-Net	0.58M	95.65%	0.96	2.37

下载: 导出CSV

表 3 不同模型的准确率随信噪比变化

SNR/dB	Accuracy
SNR/dB	ResNet-18	EfficientNetV1	MobileNetV2	ShuffleNetV2	SqueezeNet	LDSC-Net	Static	Dynamic
20	89.86	90.95	88.59	82.97	87.65	92.93	90.22	76.45
15	81.52	87.79	83.51	77.93	80.36	91.24	90.22	70.83
10	68.48	70.83	80.25	52.39	65.45	83.88	76.45	49.46
5	54.17	45.41	48.55	42.34	48.50	59.79	57.43	35.69
0	44.75	37.54	39.71	34.24	40.36	40.58	37.50	35.51

下载: 导出CSV

参考文献(25)

[1]	CRIOLLO L, MENA-ARCINIEGA C, and XING Shen. Classification, military applications, and opportunities of unmanned aerial vehicles[J]. Aviation, 2024, 28(2): 115–127. doi: 10.3846/aviation.2024.21672.
[2]	LYU Mingyang, ZHAO Yibo, HUANG Chao, et al. Unmanned aerial vehicles for search and rescue: A survey[J]. Remote Sensing, 2023, 15(13): 3266. doi: 10.3390/rs15133266.
[3]	QU Chengyi, SORBELLI F B, SINGH R, et al. Environmentally-aware and energy-efficient multi-drone coordination and networking for disaster response[J]. IEEE Transactions on Network and Service Management, 2023, 20(2): 1093–1109. doi: 10.1109/TNSM.2023.3243543.
[4]	CEVIZ O, SEN S, and SADIOGLU P. A survey of security in UAVs and FANETs: Issues, threats, analysis of attacks, and solutions[J]. IEEE Communications Surveys & Tutorials, 2025, 27(5): 3227–3265. doi: 10.1109/COMST.2024.3515051.
[5]	OMOLARA A E, ALAWIDA M, and ABIODUN O I. Drone cybersecurity issues, solutions, trend insights and future perspectives: A survey[J]. Neural Computing and Applications, 2023, 35(31): 23063–23101. doi: 10.1007/s00521-023-08857-7.
[6]	杨勇, 王雪松, 张斌. 基于时频检测与极化匹配的雷达无人机检测方法[J]. 电子与信息学报, 2021, 43(3): 509–515. doi: 10.11999/JEIT200768. YANG Yong, WANG Xuesong, and ZHANG Bin. Radar detection of unmanned aerial vehicles based on time-frequency detection and polarization matching[J]. Journal of Electronics & Information Technology, 2021, 43(3): 509–515. doi: 10.11999/JEIT200768.
[7]	江波, 屈若锟, 李彦冬, 等. 基于深度学习的无人机航拍目标检测研究综述[J]. 航空学报, 2021, 42(4): 524519. doi: 10.7527/S1000-6893.2020.24519. JIANG Bo, QU Ruokun, LI Yandong, et al. Object detection in UAV imagery based on deep learning: Review[J]. Acta Aeronautica et Astronautica Sinica, 2021, 42(4): 524519. doi: 10.7527/S1000-6893.2020.24519.
[8]	XIE Yuelei, JIANG Ping, GU Yi, et al. Dual-source detection and identification system based on UAV radio frequency signal[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 2006215. doi: 10.1109/TIM.2021.3103565.
[9]	ZHENG Haolin, GAO Ning, CAI Donghong, et al. UAV individual identification via distilled RF fingerprints-based LLM in ISAC networks[J]. IEEE Wireless Communications Letters, 2025, 14(11): 3769–3773. doi: 10.1109/LWC.2025.3603423.
[10]	MEDAIYESE O O, EZUMA M, LAUF A P, et al. Hierarchical learning framework for UAV detection and identification[J]. IEEE Journal of Radio Frequency Identification, 2022, 6: 176–188. doi: 10.1109/JRFID.2022.3157653.
[11]	唐智灵, 杨小牛, 李建东. 基于顺序统计的窄带通信辐射源指纹特征抽取方法[J]. 电子与信息学报, 2011, 33(5): 1224–1228. doi: 10.3724/SP.J.1146.2010.01058. TANG Zhiling, YANG Xiaoniu, and LI Jiandong. A novel method based on order statistics for extracting fingerprint of narrow band emitter[J]. Journal of Electronics & Information Technology, 2011, 33(5): 1224–1228. doi: 10.3724/SP.J.1146.2010.01058.
[12]	唐晓刚, 冯俊豪, 张斌权, 等. 基于射频指纹的卫星测控地面站身份识别方法[J]. 电子与信息学报, 2023, 45(7): 2554–2560. doi: 10.11999/JEIT220804. TANG Xiaogang, FENG Junhao, ZHANG Binquan, et al. Satellite telemetry track and command ground station identification method based on RF fingerprint[J]. Journal of Electronics & Information Technology, 2023, 45(7): 2554–2560. doi: 10.11999/JEIT220804.
[13]	吴一全, 童康. 基于深度学习的无人机航拍图像小目标检测研究进展[J]. 航空学报, 2025, 46(3): 30848. doi: 10.7527/S1000-6893.2024.30848. WU Yiquan and TONG Kang. Research advances on deep learning-based small object detection in UAV aerial images[J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(3): 30848. doi: 10.7527/S1000-6893.2024.30848.
[14]	蒋伊琳, 尹子茹, 宋宇. 基于卷积神经网络的低截获概率雷达信号检测算法[J]. 电子与信息学报, 2022, 44(2): 718–725. doi: 10.11999/JEIT210132. JIANG Yilin, YIN Ziru, and SONG Yu. Low probability of intercept radar signal detection algorithm based on convolutional neural networks[J]. Journal of Electronics & Information Technology, 2022, 44(2): 718–725. doi: 10.11999/JEIT210132.
[15]	AL-SA’D M F, AL-ALI A, MOHAMED A, et al. RF-based drone detection and identification using deep learning approaches: An initiative towards a large open source drone database[J]. Future Generation Computer Systems, 2019, 100: 86–97. doi: 10.1016/j.future.2019.05.007.
[16]	赵慎, 李广选, 周鲜成, 等. 多无人机射频信号CNN\| Triplet-DNN异构网络特征提取与机型识别[J]. 电子与信息学报, 2026, 48(3): 1230–1240. doi: 10.11999/JEIT250757. ZHAO Shen, LI Guangxuan, ZHOU Xiancheng, et al. Multi-UAV RF signals CNN\| Triplet-DNN heterogeneous network feature extraction and type recognition[J]. Journal of Electronics & Information Technology, 2026, 48(3): 1230–1240. doi: 10.11999/JEIT250757.
[17]	聂伟, 戴琪霏, 杨小龙, 等. 基于多维信号特征的无人机探测识别方法[J]. 电子与信息学报, 2024, 46(3): 1089–1099. doi: 10.11999/JEIT230302. NIE Wei, DAI Qifei, YANG Xiaolong, et al. Unmanned aerial vehicle detection and recognition method based on multi-dimensional signal feature[J]. Journal of Electronics & Information Technology, 2024, 46(3): 1089–1099. doi: 10.11999/JEIT230302.
[18]	聂伟, 张中洋, 杨小龙, 等. 基于梅尔倒谱系数的无人机探测与识别方法[J]. 电子与信息学报, 2025, 47(4): 1076–1084. doi: 10.11999/JEIT241111. NIE Wei, ZHANG Zhongyang, YANG Xiaolong, et al. Unmanned aerial vehicles detection and recognition method based on Mel Frequency Cepstral Coefficients[J]. Journal of Electronics & Information Technology, 2025, 47(4): 1076–1084. doi: 10.11999/JEIT241111.
[19]	田旭东, 白雪茹, 周峰. 基于稀疏自编码器的空间微动目标融合识别方法[J]. 电子与信息学报, 2023, 45(12): 4336–4344. doi: 10.11999/JEIT221163. TIAN Xudong, BAI Xueru, and ZHOU Feng. Fusion recognition of space targets with micro-motion based on a sparse auto-encoder[J]. Journal of Electronics & Information Technology, 2023, 45(12): 4336–4344. doi: 10.11999/JEIT221163.
[20]	YENDURI S, GUDAVALLI M, and C G. Optimal transport-based fusion of two-stream convolutional networks for action recognition[J]. Applied Intelligence, 2025, 55(7): 641. doi: 10.1007/s10489-025-06518-x.
[21]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 770–778. doi: 10.1109/CVPR.2016.90.
[22]	TAN Mingxing and LE Q V. EfficientNet: Rethinking model scaling for convolutional neural networks[C]. Proceedings of the 36th International Conference on Machine Learning (ICML), Long Beach, USA, 2019: 6105–6114.
[23]	HOWARD A G, ZHU Menglong, CHEN Bo, et al. MobileNets: Efficient convolutional neural networks for mobile vision applications[J]. arXiv preprint arXiv: 1704.04861, 2017. doi: 10.48550/arXiv.1704.04861.
[24]	MA Ningning, ZHANG Xiangyu, ZHENG Haitao, et al. ShuffleNet V2: Practical guidelines for efficient CNN architecture design[C]. Proceedings of the 15th European Conference on Computer Vision–ECCV 2018, Munich, Germany, 2018: 122–138. doi: 10.1007/978-3-030-01264-9_8.
[25]	IANDOLA F N, HAN Song, MOSKEWICZ M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size[C]. Proceedings of the 5th International Conference on Learning Representations, Toulon, France, 2017.