MCL-YOLO:一种细粒度特定辐射源识别方法  

MCL-YOLO: A Fine-grained Specific Emitter Identification Method

作　　者：陈海永 赵宇琦 刘坤 宿绍莹[2] CHEN Haiyong;ZHAO Yuqi;LIU Kun;SU Shaoying(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China;School of Electronic Science,National University of Defense Technology,Changsha,Hunan 410073,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300401 [2]国防科技大学电子科学学院,湖南长沙410073

出　　处：《信号处理》2023年第1期96-104,共9页Journal of Signal Processing

基　　金：国家自然科学基金项目(62073117,U21A20482);中央引导地方科技发展资金项目(206Z1701G);国家自然科学基金(面上项目)(62173124);河北省自然科学基金(F2019202305)。

摘　　要：特定辐射源识别(Specific Emitter Identification, SEI)广泛应用于电子对抗、频谱管控、无线网络安全等军民领域。针对传统SEI方法依赖先验知识、普适性差、细粒度任务难以精细识别的问题,首先,利用接收机组建采集系统,采集Wi-Fi辐射源信号的数字频谱余晖图数据,建立国内首个特定辐射源识别数据集;其次,提出关注目标局部细微特征的Mutual Channel Loss-YOLO(MCL-YOLO)网络模型,充分挖掘数字频谱余晖图三维信息,高度聚焦子类间微小差异,实现细粒度特定辐射源识别;最后,在Wi-Fi辐射源个体数据集(WFED)上进行对比实验验证。实验结果表明,MCL-YOLO在WFED上精确率(Precision, P)、召回率(Recall, R)、F1测度(F1-Score, F1)和均值平均精度(mean Average Precision, mAP)比YOLOv5s分别提高2.9%、2.5%、2.7%、1.1%。充分证明其能聚焦相似特征间的细微差异,提高模型在细粒度SEI任务中的有效性。Specific Emitter Identification(SEI) is widely used in military and civil fields, such as electronic countermeasures, spectrum control, wireless network security, etc. Traditional SEI methods rely on prior knowledge, have poor universality, and are difficult to identify fine-grained tasks. First of all, to solve the above problems, a receiver is used to set up an acquisition system to collect the digital afterglow spectrum data of Wi-Fi emitter signals, so as to establish the first specific emitter identification database in China. Next in importance, a Mutual Channel Loss-YOLO(MCL-YOLO) network model is proposed, which pays attention to the local subtle features of the target to be detected. This network can fully mine the three-dimensional information of the digital afterglow spectrum diagram, highly focus on the small differences between subclasses, and realize fine-grained specific emitter identification. In the end, the Wi-Fi Emitter Dataset(WFED) is used to verify the contrast experiment. The experimental results show that the precision(P), recall(R), F1-Score(F1) and mean Average Precision(mAP) of MCL-YOLO on the data set WFED are improved by 2. 9%, 2. 5%, 2. 7% and 1. 1% respectively compared with those of YOLOv5s. It is fully proved that MCL-YOLO network can highly focus on the subtle differences between similar features and improve the effectiveness of the model in fine-grained SEI tasks.

关 键 词：特定辐射源识别 细粒度 目标检测 YOLOv5 数字频谱余晖图 

分 类 号：TN971[电子电信—信号与信息处理] TP391.4[电子电信—信息与通信工程]