基于镜像填充谱与LA-ResNet50的超短波卫星信道分类识别算法  

Ultrashort wave satellite channel classification and recognition algorithm based on mirror filled spectrum and LA-ResNet50

作　　者：吴尚 沈雷[1] 王李军 张如栩[1] 胡鑫 WU Shang;SHEN Lei;WANG Lijun;ZHANG Ruxu;HU Xin(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China;The 36th Research Institute of China Electronics Technology Corporation,Jiaxing 314000,China;Key Laboratory of Communication Information Control and Security Technology,Jiaxing 314000,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310018 [2]中国电子科技集团公司第三十六研究所,浙江嘉兴314000 [3]通信信息控制和安全技术重点实验室,浙江嘉兴314000

出　　处：《电信科学》2023年第10期74-84,共11页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.62072121);浙江省尖兵领雁项目(No.2023C01143)。

摘　　要：针对超短波频段中存在的5 kHz信道、25 kHz信道、宽带干扰信道、窄带干扰信道和单音干扰信道的分类识别问题,提出了一种基于镜像填充谱与局部二值模式的注意力机制残差网络(LBP attention ResNet50,LA-ResNet50)的超短波信道分类识别方法,有效解决了低信噪比下卫星信道与底噪难以区分,信号信道与特征相近的干扰信道识别困难的问题。首先,所提方法对超短波的频谱进行镜像对称并填充,同时对频谱边缘进行描黑处理,构成镜像填充谱,提高不同类型信道频谱图的区分度;然后,在ResNet50中引入通道注意力机制,使网络模型关注度集中在信道上;最后,提出了基于交叉熵和局部二值模式(local binary pattern,LBP)的损失函数,提高对信号信道和干扰信道图像边缘细微纹理特征的提取效果。所提基于镜像填充谱和LA-ResNet50的方法,对比利用快速傅里叶变换(fast Fourier transform,FFT)频谱门限阈值分类的传统方法与基于镜像填充谱的YOLOv5s目标检测分类法,以及基于镜像填充谱的注意力机制残差网络(Attention-ResNet50)、Transformer网络方法,在10 dB信噪比下对超短波信道的分类识别率分别提高了19.8%、8.2%、1.8%、0.8%。In response to the classification and identification problems of 5 kHz channels,25 kHz channels,broadband interference channels,narrowband interference channels,and single tone interference channels in the ultrashort wave frequency band,a classification and identification method for ultrashort wave channels based on mirror filled spectrum and LA-ResNet50(LBP attention ResNet50)was proposed.The problem of difficulty in distinguishing between satellite channels and background noise under low signal-to-noise ratio,as well as the identification of signal channels and interference channels with similar characteristics,has been effectively solved.Firstly,the proposed method performs mirror symmetry on the ultrashort wave spectrum and fills it in,while blackening the edges of the spectrum to construct a mirror-filled spectrum,which improves the discrimination of different types of channel spectra.Then,channel attention was introduced into ResNet50 to focus the attention of the network model on the channel.Finally,a loss function based on cross entropy and local binary pattern(LBP)was proposed to improve the extraction effect of subtle texture features on signal channels and interference channels images.The proposed method based on mirror-filled spectrum and LA-ResNet50 has shown an improvement of 19.8%,8.2%,1.8%,and 0.8%in classification accuracy for ultrashort wave channels compared to the traditional method utilizing fast Fourier transform(FFT)spectrum thresholding,the YOLOv5s target detection and classification method based on mirror-filled spectrum,the Attention-ResNet50 method with attention mechanism based on mirror-filled spectrum,and the Transformer network method under a signal-to-noise ratio(SNR)of 10 dB.

关 键 词：超短波信道 注意力机制 分类识别 ResNet50 

分 类 号：TN911.73[电子电信—通信与信息系统]