基于鲁棒损失函数的标签有噪信号调制方式识别

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l₁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。

Robust Loss Functions for Signal Modulation Recognition with Noise Labels

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 l₁ norm 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.