基于数据增强和特征嵌入的自动调制识别

Automatic Modulation Recognition Based on Data Enhancement and Feature Embedding

基于端到端的深度学习模型已经被广泛应用于自动调制识别。现有的深度学习方案大多数依赖于丰富的样本分布,而大批量的标记训练集通常很难获得。提出了一种基于数据驱动和选择性核卷积神经网络(Convolutional Neural Network,CNN)的自动调制识别框架。首先开发深度密集生成式对抗网络增强5种调制信号的原始数据集;其次选择平滑伪Wigner-Ville分布作为信号的时频表示,并将注意力模块用于聚焦时频图像分类中的差异区域;最后将真实信号输入轻量级卷积神经网络进行时间相关性提取,并融合信号的时频特征完成分类。实验结果表明,所提算法提高了在低信噪比情况下的识别精度,表现出较强的鲁棒性。

The end-to-end deep learning model has been effectively applied in automatic modulation recognition(AMR).Most deep learning schemes need a large number of training samples,and a large number of labeled training sets are usually difficult to obtain.The authors propose an automatic modulation recognition framework based on data-driven and selective kernel convolution neural network(CNN).Firstly,a deep dense generative adversarial network(GAN)is developed to enhance the original data set of five modulation signals.Secondly,the smooth pseudo Wigner-Ville distribution is selected as the time-frequency representation of the signal,and the attention block is used to focus the difference region in the time-frequency image classification.Finally,the real signal is input into lightweight CNN for time correlation extraction,and the time-frequency characteristics of the signal are fused to complete the classification.Experimental results show that the proposed algorithm improves the recognition accuracy in the case of low signal-to-noise ratio and shows strong robustness.