基于鲁棒损失函数的标签有噪信号调制方式识别  

Robust Loss Functions for Signal Modulation Recognition with Noise Labels

作　　者：王晓波 尹俊平[1] 徐岩 WANG Xiaobo;YIN Junping;XU Yan(Institute of Applied Physics and Computational Mathematics,Beijing 100094,China;Department of Information and Computer Science,University of Science and Technology Beijing,Beijing 100084,China)

机构地区：[1]北京应用物理与计算数学研究所,北京100094 [2]北京科技大学信息与计算科学系,北京100083

出　　处：《计算物理》2022年第4期386-394,共9页Chinese Journal of Computational Physics

基　　金：国家自然科学基金天元基金重点项目(12026607);国家自然科学基金(12071024,12031016)资助

摘　　要：针对现实信号调制方式标注易发生错误,即训练数据集中信号调制方式标签存在噪声情形,我们选取l模损失函数及其推广形式作为对标签噪声具有鲁棒性的损失函数,结合深度卷积神经网络优良的自动特征提取能力,提出一种针对信号调制方式存在误判噪声的深度学习算法。该算法在训练数据集合标签噪声率达50%情形下,对信号调制方式的识别准确率依然保持较高水平。相反,对于采用通常的交叉熵作为损失函数的深度卷积神经网络,其已无法对信号调制方式进行分类识别。在公开的数据集上的数值实验表明,所提算法对于标签有噪信号调制方式识别具有较强的鲁棒性。In view of the fact that the labeling of signal modulation type is prone to errors in applications,that is,the underlying training data set has label noise,we propose lnorm based loss function and its extended form as robust loss function of deep convolutional neural network,which is recognized as one of the most excellent feature extraction network,to classify signal modulation types with label noisy.The algorithm achieves high accuracy even if the label noise level of training data set is up to 50%.By contrast,it is unable to predict the type of signal modulation by using usual cross entropy as the loss function of the deep convolutional neural network.Robustness of the algorithm is verified with numerical examples on public available benchmark data sets.

关 键 词：l1模损失函数 q损失函数 信号调制 有噪标签 信号识别 

分 类 号：TN911.3[电子电信—通信与信息系统]