基于度量学习和子域自适应的辐射源个体识别  

Specific Emitter Identification Based on Metric Learning and Subdomain Adaptation

作　　者：周锋 杜奕航 赵芸 乔晓强 张涛[2] ZHOU Feng;DU Yihang;ZHAO Yun;QIAO Xiaoqiang;ZHANG Tao(School of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;The 63 rd Research Institute,National University of Defense Technology,Nanjing 210007,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,南京210044 [2]国防科技大学第六十三研究所,南京210007

出　　处：《电讯技术》2025年第3期347-353,共7页Telecommunication Engineering

基　　金：国家自然科学基金资助项目(62371463)。

摘　　要：为解决辐射源个体识别中信号传输环境变化引起的数据分布不一致,导致仅接受单一分布数据集训练的网络模型识别准确率严重退化这一问题,提出结合度量学习和子域自适应的辐射源个体识别方法。该方法借鉴了领域自适应中子域自适应的思想,应用局部最大均值差异损失来缩小不同分布下相同辐射源类别之间的差异,并在其基础上加入基于欧氏距离和余弦相似度的度量学习损失,稳定迁移效果。实验表明,在同时使用了度量学习损失和子域自适应方法后,目标域识别准确率相比于未使用迁移方法提高了38.7%左右,并且模型具有良好的泛化能力。In order to solve the problem of inconsistent data distribution caused by changes in the transmission environment in specific emitter identification,which leads to serious degradation in identification accuracy of network models trained only with a single distribution data set,a method for specific emitter identification that combines metric learning and subdomain adaptation is proposed.This method draws on the idea of subdomain adaptation in domain adaptation,applies local maximum mean difference loss to reduce the differences between the same emitter identification categories under different distributions,and the metric learning loss based on Euclidean distance and cosine similarity is added to stabilize the migration effect.Experiments show that after using both the metric learning loss and the subdomain adaptation method,the target domain recognition accuracy is improved by about 38.7%compared with that of the non-transfer method,and the model has good generalization ability.

关 键 词：辐射源个体识别 度量学习 子域自适应 余弦相似度 

分 类 号：TN971[电子电信—信号与信息处理]