基于通道融合的调制信号识别方法  

Modulation signal recognition method based on channel fusion

作　　者：潘一震 韩顺利[1] 季桓勇 张博 PAN Yizhen;HAN Shunli;JI Huanyong;ZHANG Bo(The 41st Institute of China Electronic Technology Group Corporation,Qingdao 266555,China)

机构地区：[1]中国电子科技集团第四十一研究所,山东青岛266555

出　　处：《现代电子技术》2023年第12期57-62,共6页Modern Electronics Technique

摘　　要：针对现有调制方式识别存在的计算量大、网络模型复杂、识别准确率低等问题,文中提出一种基于通道融合的新型调制方式识别方法。该方法由双流卷积神经网络模块和GRU神经网络模块构成,其中双流卷积神经网络为两条并联的深度可分离卷积子网络,分别提取信号不同尺度下的空间特征,同时添加短路连接来增加特征传递与重用。将两通道提取到的特征在通道维度上进行融合,进而形成更为丰富的融合特征。将融合特征输入至GRU神经网络模块中提取信号的时序特征,提取的互补信息可使网络学习到更加全面的信号特征,从而提高调制方式识别的精度。在数据集RadioML2016.10a上进行实验,实验结果表明,所提方法的网络性能优于其他神经网络算法,信噪比在0 dB以上时识别率可达到90.8%,能够有效提高自动调制识别的准确率。In allusion to the problems of large amount of calculations,complex network model and low recognition accuracy in the existing modulation recognition methods,a method of new modulation recognition based on channel fusion is proposed in this paper.This method is composed of two⁃stream convolutional neural network module and GRU(gated cycle unit)neural network module.The two⁃stream convolutional neural network is two parallel deep separable convolutional sub⁃networks,which can extract the spatial features of the signals at different scales respectively,and add short⁃cut connections to increase feature transmission and reuse.The features extracted from the two streams are fused in the channel dimension to form richer fusion features.The fusion features are input to the GRU neural network module to extract temporal feature of the signals,and the extracted complementary information can enable it to learn more comprehensive signal features,thereby improving the recognition accuracy of the modulation mode.The experiment was conducted on dataset RadioML2016.10a.The experimental results show that the network performance of the proposed method is better than that of other neural network methods.When the signal⁃to⁃noise ratio is above 0 dB,the recognition rate can reach 90.8%,which can effectively improve the accuracy of automatic modulation recognition.

关 键 词：自动调制识别 特征提取 特征融合 特征传递 MCGNN网络模型 网络性能分析 

分 类 号：TN911.7-34[电子电信—通信与信息系统]