基于生成对抗网络的数据增强方法及应用

Generative adversarial network based data augmentation and its application in few-shot electromagnetic signal classification

摘要:

对于小样本电磁信号识别，数据增强是一种最为直观的对策。利用生成对抗网络(GAN)产生虚假信号样本，设计粗粒度和细粒度筛选机制对生成信号进行筛选，剔除质量较差的生成信号，实现训练样本集的有效扩充。为验证所提数据增强算法的有效性，在RADIOML 2016.04C数据集上进行测试。实验结果表明，本文所提方法对小样本电磁信号识别准确率有较好的提升效果。

Abstract:

For few-shot electromagnetic signal classification, data augmentation is the most intuitive strategy. In this paper, Generative Adversarial Network(GAN) is employed to generate fake signal samples. The coarse-grained and fine-grained screening mechanisms are designed to screen the generated fake signals. The generated signals with poor quality are removed and the effective expansion of training dataset is realized. In order to verify the effectiveness of the proposed data augmentation algorithm, sufficient experiments are conducted on the RADIOML 2016.04C dataset. Experimental results show that the proposed method can improve the accuracy of few-shot electromagnetic signal classification effectively.