DP-DRCnet卷积神经网络信号调制识别算法  

A Modulation Recognition Algorithm of DP-DRCnet Convolutional Neural Network

作　　者：王洋[1] 冯永新[1] 宋碧雪 田秉禾 WANG Yang;FENG Yongxin;SONG Bixue;TIAN Binghe(Key Laboratory of Information Network and Information Countermeasure Technology of Liaoning Province,Shenyang Ligong University,Shenyang 110159,Liaoning,China)

机构地区：[1]沈阳理工大学辽宁省信息网络与信息对抗重点实验室,辽宁沈阳110159

出　　处：《兵工学报》2023年第2期545-555,共11页Acta Armamentarii

基　　金：国家自然科学基金项目(61971291);中央引导地方科技发展项目(2022020128-JH6/1001);沈阳市自然科学基金项目(22-315-6-10)。

摘　　要：卷积神经网络在降低系统网络开销的同时,如何保证较高的信号调制识别准确率是目前面临的重要问题。提出一种轻量级卷积神经网络。该网络分为两路,并行提取信号的自相关和互相关特征,之后两路特征进行合并,实现不同调制方式的分类识别;该网络采用控制模型中卷积层的输入数据维度及卷积核数量的方案,实现对网络模型开销的控制。通过对多种不同的调制方式进行识别验证。实验结果表明:在信噪比为-6~12 dB条件下,其平均识别准确率可达到86.5%;与传统卷积神经网络相比,计算量降低了94.44%;与常规轻量级卷积神经网络相比,计算量降低了67.6%,该网络性能优于现有的基于轻量级卷积神经网络的调制方式识别方法。How to ensure higher signal modulation recognition accuracy while reducing system network overhead is a important problem currently faced by the convolutional neural networks.To this end, a lightweight convolutional neural network is proposed.This networkis split into two paths to parallelly extract auto-correlation and cross-correlation features of signal.Then. features from these two paths are combined so that the network can ultimately achieve classification and recognition with different modulation modes. In addition, the overhead of the network is controlled by adopting the scheme of controlling the input data dimension of the convolution layer and the number of convolution cores in the model.The recognition verification of different modulation modes is performed.The experimental result shows that: the average recognition accuracy reaches 86.5% when the signal-to-noise ratio is in the range of-6~12 dB;compared with the conventional convolutional neural network, the computational load is reduced by 94.44%;compared with the regular lightweight convolutional neural network, the computational load is reduced by 67.6%.The performance of the proposed network is better than the existing modulation recognition methods based on lightweight convolutional neural network.

关 键 词：调制识别 卷积神经网络 特征提取 深度学习 

分 类 号：TN97[电子电信—信号与信息处理]