基于改进半监督生成对抗网络的少量标签轴承智能诊断方法

Intelligent diagnosis method for bearings with few labelled samples based on an improved semi-supervised learning-based generative adversarial network

轴承在线监测大数据主要是由大量无标签数据和少量有标签数据组成的。已有的智能诊断方法多依赖于大量有标签数据的监督学习。针对此问题,提出一种改进半监督生成对抗网络,利用生成器与分类器的对抗学习增强分类器的辨识能力,提出了增强特征匹配算法,借助分类器的深层特征优化网络损失值计算,进而提高网络收敛速度,结合半监督学习使用少量有标签数据进一步提升分类器的学习能力,最终实现对无标签数据的正确归类。使用四组轴承试验数据的迁移学习验证和对比了改进深度网络对不同轴承、工况、故障生成模式、故障程度的辨识能力,结果表明该网络具有更高的分类准确率和更快的收敛速度。

The big data of bearings obtained in online monitoring are mainly composed of a large amount of data without labels and a small amount of data with labels. But, successful applications of many popular intelligent fault diagnosis methods rely on the supervised learning of large scale of labeled data. To solve this problem, an improved semi-supervised learning-based generative adversarial network(ISSL-GAN) was proposed. In the method, the classifier was endowed with more powerful classification capability through the adversarial learning mode between the generator and classifier. Furthermore, an enhanced feature matching algorithm was proposed to extract deep features from the intermediate layers of the whole network and join the loss function to speed up the convergence of the training process. Meanwhile, the semi-supervised learning was applied to make full use of few labeled data to improve the learning efficiency of the classifier. After training with plenty of unlabeled data and few labeled data, the obtained classifier can accurately discriminate the classes of analyzed data. The proposed network was applied to four experimental data sets, in each of them, through transfer learning from one bearing to another one, the classification ability for different bearings, working conditions, fault generation methods, and damage levels was verified and the results were compared with those by other deep networks. The results indicate that the proposed ISSL-GAN has much higher accuracy on bearing fault diagnosis and has faster convergence speed during training.