P-Net框架下的小样本雷达信号识别算法

Small sample radar signal recognition algorithm under the P-Net framework

针对雷达信号由于缺少先验知识难以形成信号模板或识别网络的识别难题,本文重新审视了战争雷达的应用场景,并提出一种基于原型网络的雷达信号识别算法。原型网络作为一种弱监督学习模型,已被证明对于小样本数据的分类拟合具有积极作用。本文利用卷积神经网络将现有的充足样本信号时频信息映射到样本空间,获取各充足类样本原型,以样本点到原型的距离为损失函数,使得类间分散、类内聚合,得到泛化的训练模型。实验结果表明:小样本信号在较少先验信息的支持和微调下,通过泛化的模型能准确识别信号类别。在多种小样本信号并存的条件下,-2 dB时识别率可达90%以上。

To address the difficulty of forming a signal template or a recognition network in radar signal recognition due to the lack of prior knowledge,this study reexamines the application scenarios of war radar and proposes a radar signal recognition algorithm based on a prototype network.As a weakly supervised learning model,the prototype network has been shown to have a positive effect on the classification and fitting of small sample data.Specifically,this study uses the convolutional neural network to map the existing sufficient time-frequency information of the sample signals onto the sample space and obtains the prototype of each sufficient class.The distance from the sample point to the prototype is used as the loss function to achieve interclass and intraclass dispersion and obtain a generalized training model.The experimental results show that,with the support and fine-tuning of small sample signals,the generalized model can accurately identify the signal category.Under the condition that a variety of small sample signals coexist,the recognition rate can reach more than 90%at-2 dB.