| 摘要: |
| 针对低信噪比条件下自动调制识别准确率不高的问题，提出了通道门控Res2Net卷积神经网络自动调制识别模型。该模型主要由二维卷积神经（Two-dimensional Convolutional Neural Network,2D-CNN）网络、多尺度残差网络（Residual 2-network,Res2Net）、压缩与激励网络（Squeeze-and-Excitation Network,SENet）和长短期记忆（Long Short-Term Memory,LSTM）网络组成，通过卷积从原始I/Q数据中提取多尺度特征，结合门控机制对特征通道进行权重调整，并利用LSTM对卷积所得特征进行序列建模，确保数据特征被有效挖掘，从而提升自动调制识别的准确率。在基准数据集RML2016.10a下的调制识别实验表明，所提模型在信噪比为12 dB时识别精度为92.68%，在信噪比2 dB以上时平均识别精度大于91%，较经典CLDNN模型、LSTM模型和同类型PET-CGDNN模型、CGDNet模型能取得更高的调制类型识别准确率。 |
| 关键词: 自动调制识别 卷积神经网络 压缩与激励网络 多尺度残差网络 长短期记忆网络 |
| DOI：10.20079/j.issn.1001-893x.230829006 |
|
| 基金项目: |
|
| Channel Gated Res2Net Convolutional Neural Network for Automatic Modulation Recognition |
| CHEN Hao,GUO Wenpu,KANG Kai |
| ((College of Combat Support,Rocket Force University of Engineering,Xi揳n 710025,China)) |
| Abstract: |
| In response to the problem of low accuracy in automatic modulation recognition under low signal-to-noise ratio(SNR) conditions,the authors propose a channel gated residual 2-network(Res2Net) convolutional neural network(CNN) model.The model mainly consists of two-dimensional CNN(2D-CNN),multi-scale Res2Net,squeeze-and-excitation network(SENet) and long short-term memory(LSTM) network,which extracts multi-scale features from raw I/Q data through convolution,adjusts the weight of feature channels through gating mechanism,and uses LSTM to model the sequence of convolutional features to ensure effective data feature mining,thereby improving the accuracy of automatic modulation recognition.The modulation recognition experiment on the benchmark dataset RML2016.10a shows that the recognition accuracy of the proposed model is 92.68% at 12 dB SNR,and the average recognition accuracy is above 91% when the SNR is greater than 2 dB.Compared with classical CLDNN model,LSTM model,similar PET-CGDNN model and CGDNet model,the proposed model can achieve higher modulation type recognition accuracy. |
| Key words: automatic modulation recognition convolutional neural network squeeze-and-excitation network multi-scale residual network long short-term memory network |
