基于深度残差收缩注意力网络的雷达信号识别方法  

Radar signal recognition method based on deep residual shrinkage attention network

作　　者：曹鹏宇 杨承志[1] 陈泽盛 王露[2] 石礼盟 CAO Pengyu;YANG Chengzhi;CHEN Zesheng;WANG Lu;SHI Limeng(School of Air Operations and Services,Aviation University of Air Force,Changchun 130022,China;School of Aeronautical Foundation,Aviation University of Air Force,Changchun 130022,China;Unit 93671 of the PLA,Nanyang 474350,China)

机构地区：[1]空军航空大学航空作战勤务学院,吉林长春130022 [2]空军航空大学航空基础学院,吉林长春130022 [3]中国人民解放军93671部队,河南南阳474350

出　　处：《系统工程与电子技术》2023年第3期717-725,共9页Systems Engineering and Electronics

摘　　要：针对低信噪比条件下雷达信号识别率低,以及分类网络不具备识别样本库新添加信号类型的局限,提出了一种基于深度残差收缩注意力网络的雷达信号识别方法。通过网络将一维雷达信号映射到32维向量空间。网络中的残差连接能有效强化特征的传播能力,解决网络过深无法训练的问题;注意力机制的引入,不仅构建掩码支路充当主干支路的特征选择器,还能够帮助网络自适应地选择合适的阈值进行软阈值化,从而减少网络中噪声或者冗余信息的影响,提高网络对噪声的鲁棒性。训练过程中排序表损失(ranked list loss,RLL)和分类损失函数共同指导网络训练。RLL能够有效克服传统度量学习损失函数忽略类内特征的问题,分类损失函数能够弥补度量损失优化下对样本整体分布不敏感的问题。实验表明,该方法在提高低信噪比雷达信号识别准确率的同时仍具有识别样本库新添加信号类型的能力。Aiming at the low recognition rate of radar signals under the condition of low signal-to-noise ratio,and the classification network has the limitation of identifying the newly added signal types in the sample library,a radar signal recognition method based on the deep residual shrinking attention network is proposed.The one-dimensional radar signal is mapped to a 32-dimensional vector space through the network.The residual connection in the network can effectively strengthen the dissemination ability of features and solve the problem that the network is too deep to be trained;the introduction of the attention mechanism not only constructs the mask branch to act as the feature selector of the main branch,but also helps the network adaptively select the appropriate threshold for soft thresholding,so as to reduce the influence of noise or redundant information in the network and improve the robustness of the network to noise.During the training process,ranked list loss(RLL)and the classification loss function jointly guide the network training.RLL can effectively overcome the problem of traditional metric learning loss function ignoring features within the class,and the classification loss function can make up for the problem of insensitivity to the overall distribution of the sample under metric loss optimization.Experiments show that this method can not only improve the recognition accuracy of low-signal-to-noise ratio radar signals,but also has the ability to identify newly added signal types in the sample library.

关 键 词：雷达信号识别 深层残差收缩注意力网络 软阈值化 注意力机制 损失函数 

分 类 号：TN974[电子电信—信号与信息处理]