邹波蓉, 李姗姗, 叶沛然,等.基于多通道残差网络与注意力机制协作的调制分类[J].河南理工大学学报（自然科学版）,2023,42(1):160-167.

ZOU B R, LI S S, YE P R,et al.Modulation classification based on multi-channel residual network and attention mechanism[J].Journal of Henan Polytechnic University(Natural Science) ,2023,42(1):160-167.

基于多通道残差网络与注意力机制协作的调制分类

邹波蓉1, 李姗姗2, 叶沛然3, 侯庆华1, 武会斌1

1.河南理工大学 物理与电子信息学院，河南 焦作  454000；2.商丘工学院 信息与电子信息工程学院，河南 商丘  476000；3.郑州经贸学院 土木建筑学院，河南 郑州 451191

摘要:为了更加充分地提取调制信号特征，构建一种多通道残差网络与注意力机制协作的调制分类方法。首先设计一个各通道均不相同的多通道结构，确保提取的信号特征更加多样；其次，将每个通道提取的特征利用concatenate层进行融合，增强描述信号特征；之后，结合残差网络的优势，显著增加网络深度，捕获更具代表性的特征，同时缓解深层网络带来的梯度消失问题；最后，为了使提取的特征更加易于分类，引入注意力层，对提取特征重新校准，以捕获更加关键的特征，增加信号分类准确率。在公共数据集RadioML 2016.10 b上进行实验。仿真结果表明，该网络的分类性能优于许多文献中的分类器，当信噪比14 dB时，分类精度达到93.23%，证明了此网络的可行性与有效性。

关键词:调制分类;深度学习;卷积神经网络;残差网络;注意力机制

doi:10.16186/j.cnki.1673-9787.2021120081

基金项目:国家自然科学基金资助项目（11805052）；河南省科技攻关项目（222102210247）

收稿日期:2021/12/22

修回日期:2022/01/17

出版日期:2023/01/25

Modulation classification based on multi-channel residual network and attention mechanism

ZOU Borong1, LI Shanshan2, YE Peiran3, HOU Qinghua1, WU Huibin1

1.School of Physics and Electronic Information，Henan Polytechnic University，Jiaozuo  454000，Henan，China；2.School of Information and Electronic Engineering，Shangqiu Institute of Technology，Shangqiu  476000，Henan，China；3.College of Civil Engineering and Architecture，Zhengzhou University of Economics and Business，Zhengzhou  451191，Henan，China

Abstract:In order to more fully extract the characteristics of modulated signals，a multi-channel modulation classification was constructed with the collaboration of residual network and attention mechanism.Firstly，a multi-channel structure with different channels was designed，which ensured that the extracted signal features were more diverse.Secondly，the extracted features of each channel were fused by concatenate layer，which increased the features describing the signal.Then，combining the advantages of residual network，the depth of the network was significantly increased，more representative features were captured，while the problem of gradient disappearance caused by deep network was alleviated.Then，in order to make the above extracted features easier to classify，attention layer was introduced to recalibrate the extracted features to capture more critical features and increase the accuracy of signal classification.An experiment was conducted on the public data set RadioML 2016.10b，and the simulation results showed that the classification performance of the proposed network was better than many classifiers in the literature，with a classification accuracy of 93.23% when the signal-to-noise ratio was equal to 14 dB，this demonstrated its feasibility and effectiveness.

Key words:modulation classification;deep learning;convolutional neural network;residual network;attention mechanism

 019_2021120081_邹波蓉_H.pdf