基于深度学习网络融合的自动调制分类方法
Automatic Modulation Classification Method Based on Deep Learning Network Fusion
-
摘要: 基于深度学习网络的自动调制分类（Automatic Modulation Classification， AMC）方法虽然对大多数通信调制信号能够取得满意的分类效果，但对WBFM（Wide Band Frequency Modulation）信号和MQAM（Multiple Quadrature Amplitude Modulation）信号的分类并不理想。针对WBFM信号误判的问题，使用判决法来筛选WBFM信号；考虑到信号样本不平衡的情况，引入数据增强方法扩充筛选后的WBFM信号。针对MQAM信号混淆的问题，利用分数阶傅里叶变换（Fractional Fourier Transform， FRFT）获取时频维度更多的特征信息。在此基础上，提出一种基于特征金字塔网络和长短时记忆网络并联的多通道特征融合网络（Multi-channel Feature Fusion， MFF）来提取信号的深层特征和浅层特征进行分类。实验结果表明：本文所提方法在一定程度上能够解决WBFM信号的误判问题和MQAM信号的混淆问题；与CNN（Convolutional Neural Network）、ResNet（Residual Network）、LSTM（Long Short Term Memory）、CLDNN（Convolutional Long Short-term Deep Neural Network）网络相比，所提网络具有更高的分类准确率。Abstract: Although the automatic modulation classification （AMC） method based on deep learning network can achieve satisfactory classification results for most communication modulation signals， it is not ideal for the classification of WBFM signals and MQAM signals. Aiming at the problem of misjudgment of WBFM signal， the decision method is used to screen it. Considering the imbalance of signal samples，a data enhancement method is introduced to expand the screened WBFM signals. To solve the problem of confusion of MQAM signals， fractional Fourier transform （FRFT） is adopted to obtain more feature information of time-frequency dimension. On this basis，a multi-channel feature fusion network that is composed of feature pyramid network and long short-term memory network is proposed to extract deep and shallow features for classification. The experiment results show that the proposed method can solve the problem of misjudgment of WBFM signals and the confusion of MQAM signals to a certain extend. Compared with CNN， ResNet， LSTM and CLDNN networks， the proposed network has higher classification accuracy.
下载: