基于CWD和残差收缩网络的调制方式识别方法

针对低信噪比时莱斯信道下特征提取准确性难以保证、识别准确率偏低等问题, 提出一种基于Choi-Williams分布(Choi-Williams distribution, CWD)和深度残差收缩网络(deep residual shrinkage network, DRSN)的通信辐射源信号调制方式识别方法。利用CWD将时域复信号转换为二维时频矩阵, 对深度残差网络添加软阈值化得到DRSN, 将时频矩阵样本用于对DRSN的训练, 最终构建不同信噪比下的调制方式识别网络。仿真实验表明, 基于RadioML2016.10a数据集, 利用部分先验信息的情况下, 该分类识别方法具有较高的识别准确率和噪声鲁棒性。在0 dB时, 对11类信号的总体识别准确率达到了89.95%;在2 dB以上时, 总体识别准确率均超过91%, 优于其他深度学习识别方法。

Modulation recognition method based on CWD and residual shrinkage network

Aiming at the problems of difficulty in guaranteeing the accuracy of feature extraction and poor recognition performance at low signal to noise ratio under the Rice channel, a modulation recognition method for communication radiator signal based on Choi-Williams distribution (CWD) and deep residual shrinkage network (DRSN) is proposed. In this work, CWD is used to convert the time-domain complex signals into two-dimensional time-frequency matrices firstly. Meanwhile, the soft thresholding is added to the deep residual networks (ResNets) to obtain the DRSN. Subsequently, the time-frequency matrices are used to train the DRSN. Modulation recognition network under different signal to noise ratios are finally constructed. Simulation experiments based on the RadioML2016.10a data show that the recognition network constructed has high accuracy and strong robustness to noise by utilizing partial prior information. At 0 dB, the overall recognition accuracy of the 11 types of signals reaches 89.95%. Above 2 dB, the overall recognition accuracy of the 11 types of signals exceeds 91%, which is better than other modulation recognition methods based on deep learning.