基于YOLOv8n-PEM的低信噪比无人机射频信号识别  

Low SNR Drone Radio Frequency Signal Identification Based on YOLOv8n-PEM

作　　者：刘坤 孔令轩 晏行伟 LIU Kun;KONG Lingxuan;YAN Xingwei(School of Artificial Intelligence,Hebei University of technology,Tianjin 300000,China;School of Electronic Science,National University of Defense Technology,Changsha 410000,China;Tianjin Advanced Technology Research Institute,Tianjin 300000,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300000 [2]国防科技大学电子科学学院,长沙410000 [3]天津先进技术研究院,天津300000

出　　处：《电光与控制》2025年第4期96-103,共8页Electronics Optics & Control

基　　金：国家自然科学基金(62173124,62101563);河北省自然科学基金(F2022202064)。

摘　　要：针对当前无人机射频信号识别模型在低信噪比条件下识别准确率较低,且不支持检测信号持续时间、带宽等关键参数的问题,提出了基于YOLOv8n-PEM目标检测模型的无人机射频信号识别方法。该方法首先将原始无人机射频信号做基于离散小波变换的降采样处理,之后进行短时傅里叶变换提取时频特征,最后利用YOLOv8n-PEM模型完成信号的识别和参数估计。在模型方面,基于部分卷积设计了CPF模块增强对高级时频特征的提取能力,提升模型的鲁棒性,同时引入EMA机制抑制背景噪声对模型推理的干扰。实验结果表明,YOLOv8n-PEM模型在-20~-10 dB低信噪比条件下mAP达到了96.08%,FPS为107帧/s,模型参数量较基础模型减少了38%,具有实际部署价值。Aiming at the problem that the current drone RF signal identification models have low identification accuracy under the condition of low SNR,and do not support the key parameters such as detection signal duration and bandwidth,this paper proposes a drone radio frequency signal identification method based on YOLOv8n-PEM target detection model.Firstly,the original drone RF signal is downsampled based on discrete wavelet transform,and then the time-frequency characteristics are extracted by short-time Fourier transform.Finally,the signal identification and parameter estimation are completed by using YOLOv8n-PEM model.In terms of the model,the CPF module is designed based on partial convolution to enhance the extraction ability of advanced time-frequency features and improve the robustness of the model.At the same time,the EMA mechanism is introduced to suppress the interference of background noise on the model reasoning.The experimental results show that the YOLOv8n-PEM model has an mAP of 96.08% and an FPS of 107 frames per second under low SNR conditions of-20 to-10 dB,the model parameters are reduced by 38% compared with the baseline model,indicating its value for practical deployment.

关 键 词：无人机 射频信号识别 YOLOv8 时频图 部分卷积 EMA机制 

分 类 号：TN911.7[电子电信—通信与信息系统]