李辉, 王悦悦, 魏坡,等.小样本条件下雷达信号的生成与轻量化识别[J].河南理工大学学报（自然科学版）,2024,43(5):142-151.

LI H, WANG Y Y, WEI P, et al.Radar signal generation and lightweight identification under the condition of small samples[J].Journal of Henan Polytechnic University(Natural Science) ,2024,43(5):142-151.

小样本条件下雷达信号的生成与轻量化识别

李辉, 王悦悦, 魏坡, 邹波蓉, 王伟东

河南理工大学 物理与电子信息学院，河南 焦作 454000

摘要: 目的 针对目前深度学习方法在雷达信号识别中需要海量数据且网络复杂、计算量大、设备要求高等问题，  方法 提出一种联合改进CycleGAN和MobileNetV3-Small轻量化卷积神经网络的雷达信号识别算法。首先，选取8种常见的雷达信号类型构建时域序列，为了更好保留时频特征，在信号预处理阶段将其通过崔-威廉斯分布形成图像数据集，在数据集扩充阶段将图像数据集作为CycleGAN迁移网络的输入，约束指导目标图像的生成，以解决样本不足的问题；然后，在CycleGAN的生成器中引入U-Net结构和残差密集块并更改判别器的判别方式和损失函数，以解决数据集扩增过程中的特征模糊和梯度消失等问题；最后，在信号识别阶段，通过构建具有代表性的MobileNetV3-Small轻量化网络，完成识别验证任务。  结果 图像生成网络CycleGAN的图像评价指标PSNR为39.74 dB，SSIM为0.95；MobileNet-Small信号识别网络模型迭代训练100次的参数量为1 538 942，总运行时间为2 152 s，FLOPs为127 351 188，准确率为99.30%。 结论 本文算法生成的图像与真实样本相似度高、失真度小，在不以牺牲准确率为代价的前提下识别速度有很大提升，有效实现了小样本条件下雷达信号的高精度识别。

关键词:雷达信号识别;崔-威廉斯分布;残差密集块;CycleGAN;MobileNetV3-Small

doi:10.16186/j.cnki.1673-9787.2022110002

基金项目:国家自然科学基金资助项目（62101176）

收稿日期:2022/11/01

修回日期:2023/04/19

出版日期:2024/07/31

Radar signal generation and lightweight identification under the condition of small samples

LI Hui, WANG Yueyue, WEI Po, ZOU Borong, WANG Weidong

School of Physics and Electronic Information Engineering，Henan Polytechnic University，Jiaozuo 454000，Henan，China

Abstract: Objectives In order to study the problems that the current deep learning method requires massive data，complex network，large amount of computation，and high equipment requirements in radar signal recognition， Methods a radar signal generation and recognition algorithm jointly improving CycleGAN and MobileNetV3-Small was proposed.Firstly，eight common radar signal types were selected and combined to construct the time-domain sequence.In order to better preserve the time-frequency features，the image dataset was formed through the Choi-Williams distribution in the signal pre-processing stage.In the dataset expansion stage，the image dataset was used as the input of the CycleGAN migration network and constrained to guide the generation of target images to solve the problem of insufficient samples.Secondly，the U-Net structure and residual dense blocks were introduced into the generator of CycleGAN and the discriminator discriminant and loss function were changed to solve the problems of feature blurring and gradient disappearance during the dataset augmentation.Finally，in the signal recognition stage，a representative MobileNetV3-Small lightweight network was constructed to complete the recognition verification task. Results The image evaluation index of CycleGAN of the image generation network was 39.74 dB for PSNR and 0.95 for SSIM，the number of parameters for 100 iteration training of the MobileNet-Small signal recognition network model was 1 538 942，and the total running time was 2 152 s.The FLOPs was 127 351 188，and the accuracy rate was 99.30%.  Conclusions The image generated by the proposed algorithm had high similarity and small distortion with the real sample，and the recognition speed was greatly improved without sacrificing the accuracy rate，which effectively realized the high accuracy recognition of radar signals under small sample conditions.

Key words:radar signal recognition;Choi-Williams distribution;residual dense block;CycleGAN;MobileNetV3-Small