自监督双流融合的小样本雷达辐射源识别方法  

Few-shot Radar Emitter Identification Based on Self-supervised Dual-stream Fusion

作　　者：郑博元 丛迅超 胡超 陈杰梅 ZHENG Boyuan;CONG Xunchao;HU Chao;CHEN Jiemei(Southwest China Institute of Electronic Technology,Chengdu 610036,China)

机构地区：[1]中国西南电子技术研究所,成都610036

出　　处：《电讯技术》2023年第9期1340-1347,共8页Telecommunication Engineering

摘　　要：针对实际场景中辐射源数据稀缺造成的小样本问题,提出了一种基于自监督和双流融合的小样本雷达辐射源识别方法。首先利用高斯分布噪声、莱斯多径衰落、设计时钟偏移信号等减损方法,基于有限数量的真实样本构建类均衡辐射源信号样本集。基于增强数据集,提出一种信号时间序列与时频图的双流特征融合模型。采用对比学习方法构建双流特征融合模型的自监督上游任务,以提升在有限标签数据情况下信号多域特征的表征能力与泛化能力。实验结果证明,该方法在小样本条件下能够有效地实现较好的辐射源类型识别能力,在目标域每个类别100个样本限制下,识别精度达到97.1%,与传统一维特征方法和基于长短期记忆(Long Short Term Memory,LSTM)的方法相比均有较大提升。The few-shot radar emitter identification method based on self-supervision and dual-stream fusion is proposed to address the problem of sparse radiation source data in practical scenarios.Firstly,loss reduction methods such as Gaussian noise,Rician multipath fading,and designing clock offset signals are used to construct an augmented dataset of radiation source signals based on a limited number of actual sampling samples.A contrastive learning dual-stream fusion model that can use both signal time series and time-frequency diagrams is proposed to enhance the representation and generalization ability.Experimental results demonstrate that the proposed method can effectively achieve good emitter type identification ability under few-shot learning conditions.When there are 100 samples of each class in the target domain,the proposed mehod achieves 97.1%recognition accuracy,which is significantly improved compared with that of traditional one-dimensional feature methods and Long Short Term Memory(LSTM)based methods.

关 键 词：雷达辐射源识别 自监督学习 小样本学习 双流特征融合 

分 类 号：TN971[电子电信—信号与信息处理]