基于MDSCLDNN-HAN的调制识别算法  

Modulation Recognition Algorithm Based on MDSCLDNN-HAN

作　　者：李天宇 侯进[1,3] 李昀喆 郝彦超 LI Tianyu;HOU Jin;LI Yunzhe;HAO Yanchao(IPSOM Lab,School of Information Science and Technology,Southwest Jiaotong University,Chengdu 611756,China;Tangshan Institute,Southwest Jiaotong University,Tangshan 063000,China;National Engineering Laboratory of Integrated Transportation Big Data Application Technology,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学信息科学与技术学院智能感知智慧运维实验室,四川成都611756 [2]西南交通大学唐山研究院,河北唐山063000 [3]西南交通大学综合交通大数据应用技术国家工程实验室,四川成都611756

出　　处：《无线电工程》2022年第9期1525-1532,共8页Radio Engineering

基　　金：国家重点基础研究发展计划(2014CB845800);四川省科技计划项目(2020SYSY0016)。

摘　　要：针对基于深度学习的调制识别模型存在模型参数多、计算量大等问题,使用深度可分离卷积和注意力机制,提出了一种新型多通道特征融合的神经网络模型。在数据集RadioML2016.10a和RadioML2016.10b上进行实验,验证结果表明,信噪比在0 dB以上时,所提算法模型对2个数据集的识别准确率分别为92.9%和93.1%,识别准确率优于现有模型,同时参数量减少65.7%,计算量减少76.6%。For the problems that the current deep learning-based modulation recognition models that had many model parameters and large amount of calculation,a new type of multi-channel feature fusion neural network model is proposed by using the depthwise separable convolution and attention mechanisms.The experiments were performed on the data sets of RadioML2016.10 a and RadioML2016.10 b.The experimental results show that when the signal-to-noise ratio is above 0 dB,the proposed algorithm model has a recognition accuracy of 92.9%and 93.1%for the two data sets,respectively.The recognition accuracy rate is better than the existing model,while the parameter amount is reduced by 65.7%,and the calculation amount is reduced by 76.6%,which has practical application value.

关 键 词：调制识别 深度学习 深度可分离卷积 幅度和相位 注意力 

分 类 号：TP927.2[自动化与计算机技术]