Application of Generative Adversarial Networksbased Semi-supervised Learning For Unlabeled Bearing Diagnosis

In many industrial processes, most of monitoring data are significant unlabeled data with limited accidental information so as to ensure long-term stable and safe operations of the monitored machines. Many classification methods based on the supervised learning usually lead to unsatisfied results, for example, the misclassification of data from minority categories as dominant categories. Compared with the unsupervised learning, the semi-supervised learning fully utilizes few labeled data and is one of the most promising learning algorithms to improve the labeling accuracy. After combining with the generative adversarial networks (GANs) and feature matching, the GANsbased semi-supervised learning (SSL) takes better learning performance, and thus is introduced for intelligent diagnosis of bearings in this paper. For the condition monitoring of bearings in rotating machinery, there are few historical data with labels and a large of amount monitoring data without labels. To process such bearing vibration data, the discriminator in the GANs is used as a classifier, and the generator with feature matching output expected features of an intermediate layer in the discriminator. Both of them are trained at the same time by using an adversarial learning procedure. In experiments, the GAN-SSL method is trained and verified on different bearings. Furthermore, the performances on the classification accuracy are also compared between the GAN-SSL and the commonly used convolutional neural network. The experimental results demonstrate that the GAN-SSL can efficiently learn the information from some bearings and predict states of other bearings with higher accuracy, and thus it can be further extended to intelligent knowledge transfer and online diagnosis.