基于星座图和密集连接网络的QAM信号识别  

Modulation Classification Based on Constellation Diagrams and DenseNet for QAM Signals

作　　者：葛战 李兵[1] 孙磊 蒋鸿宇[1] 周劼[1] GE Zhan;LI Bing;SUN Lei;JIANG Hongyu;ZHOU Jie(Institute of Electronic Engineering,China Academy of Engineering Physics,Mianyang 621000,China)

机构地区：[1]中国工程物理研究院电子工程研究所,四川绵阳621000

出　　处：《电子信息对抗技术》2023年第1期43-48,共6页Electronic Information Warfare Technology

基　　金：国家自然科学基金委员会与中国工程物理研究院联合基金(NSAF)资助项目(U153010137)。

摘　　要：利用深度神经网络对图像数据的显著学习能力,提出了一种基于通信信号星座图和密集连接网络(DenseNet)的正交幅度调制(Quadrature Amplitude Modulation,QAM)信号调制分类算法。算法首先对接收信号进行预处理,获取训练所需的星座图数据集,然后采用DenseNet对其进行训练学习,进而实现调制分类。同时在DenseNet网络中引入通道注意力机制,进一步增强特征的学习能力,提升分类性能。对不同信号长度、不同神经网络以及存在频偏和相偏估计误差等场景进行了多组实验。仿真结果表明:DenseNet相比卷积神经网络(Convolutional Neural Network,CNN)和残差网络ResNet-20能够有效提升分类准确率;相比基于累积量、累积分布及原始IQ(In-phase and Quadrature)数据等现有方法也均取得了较优的识别性能。Taking advantage of the excellent learning ability of deep neural networks for images,a modulation classification algorithm based on constellation diagrams and DenseNet is proposed for quadrature amplitude modulation(QAM)signals.The received signals are first converted into constellation diagrams.Then,DenseNet is designed to learn from the images for further modulation classification.In addition,the channel attention mechanism is introduced into the DenseNet network,which further enhances the ability to extract features from constellation diagrams and improves the classification performance.Moreover,various experiments are carried out for scenarios with different signal lengths,different neural networks,frequency offsets,and phase offsets.Simulation results show that DenseNet can effectively improve the classification accuracy compared with convolutional neural network(CNN)and ResNet-20 networks.It also achieves better classification performance than the existing methods based on cumulants,cumulative distribution function,and IQ(In-phase and Quadrature)data.

关 键 词：密集连接网络 调制分类 星座图 注意力机制 

分 类 号：TN911[电子电信—通信与信息系统]