| 摘要: |
| 针对基于深度学习的调制识别方法存在的未利用原始信号顺序信息、识别率低、参数量大的问题，提出一种基于时空卷积网络（Spatiotemporal Convolutional Network,SCN）的调制识别算法。为防止信号的顺序信息的丢失，该网络先提取信号的时域特征，再提取信号的空间特征，其中时域特征提取采用时序卷积网络(Temporal Convolutional Network,TCN)结构，空间特征提取采用二维卷积神经网络(Two-Dimensional Convolution Neural Network,2D-CNN)，最后的分类识别采用全局平均池化(Global Average Pooling,GAP) 替代展平(Flatten)层。由于TCN中因果膨胀卷积和GAP的应用使网络高识别率的同时参数大幅减少。在未经预处理的IQ信号调制识别中，与传统的CNN2、ResNet、DenseNet、CLDNN和LSTM2相比，参数量最少，平均识别精度提升4.9%~16.5%。 |
| 关键词: 通信信号 调制识别 深度学习 时域特征 空间特征 全局平均池化 |
| DOI：10.20079/j.issn.1001-893x.240116003 |
|
| 基金项目:重庆市自然科学基金创新发展联合基金（中国星网）（CSTB2023NSCQ-LZX0114） |
|
| Modulation Recognition of Communication Signals Based on Spatiotemporal Convolutional Network |
| CHEN Fatang,LIU Ze,FAN Zijian |
| (a.School of Communication and Information Engineering;b.School of Software Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China) |
| Abstract: |
| For the problems that modulation recognition methods based on deep learning do not utilize the original signal sequence information,the recognition rate is low,and the number of parameters is large,a modulation recognition algorithm based on Spatiotemporal Convolutional Network(SCN) is proposed.In order to prevent the loss of signal sequence information,the network first extracts the temporal features of the signal,and then extracts the spatial features of the signal.The temporal features are extracted using the Temporal Convolutional Network(TCN) structure.Two-Dimensional Convolution Neural Network(2D-CNN) is used to extract spatial features.In the final classification,Global Average Pooling(GAP) is used to replace the Flatten layer.Due to the application of causal expansion convolution and GAP in TCN,the simultaneous parameters of high recognition rate are greatly reduced.Compared with that of the traditional CNN2,ResNet,DenseNet,CLDNN and LSTM2,the IQ signal modulation recognition without preprocessing has the lowest number of parameters,and the average recognition accuracy is improved by 4.9%~16.5%. |
| Key words: communication signal modulation recognition deep learning time domain characteristics spatial characteristics global average pooling |
