A multi-stage semi-supervised learning approach for intelligent fault diagnosis of rolling bearing using data augmentation and metric learning. (1st January 2021)

• Record Type: Journal Article
• Authors: Yu, Kun; Lin, Tian Ran; Ma, Hui; Li, Xiang; Li, Xu
• Is Part Of: Mechanical systems and signal processing. Volume 146(2021)
• Publication Date: 2021-01-01
• Subjects: Rolling bearing -- Intelligent fault diagnosis -- Semi-supervised learning -- Data augmentation -- K-means -- Kullback-Leibler divergence; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2020.107043
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A SSL method for intelligent fault diagnosis of rolling bearing is proposed. Data augmentation and metric learning are the main elements of the proposed method. A multi-stage strategy is formulated to improve the identification ability. The proposed method achieves excellent performance on two case studies. Abstract: Limited condition monitoring data are recorded with label information in practice, which make the fault identification task more challenging. A semi-supervised learning (SSL) approach can be employed to increase the identification performance of the classifiers under such situation. In this study, a three-stage SSL approach using data augmentation (DA) and metric learning is proposed for an intelligent bearing fault diagnosis under limited labeled data. In the first stage, a DA method comprising seven DA strategies is presented to expand the feature space for the limited labeled samples under each healthy conditions. An optimization objective combining a cross entropy loss and a triplet loss is adopted to enlarge the margin between the feature distributions of limited labeled samples under different healthy conditions. In the second stage, a K-means technique is employed to acquire the cluster centers for the limited labeled samples under different healthy conditions. In the third stage, the label information for the unlabeled samples is first estimated according to the membership between the feature distributions of the unlabeled samples and the various cluster centers for original labeled samples and then a Kullback-Leibler divergence loss is introduced to minimize the discrepancy between feature distributions for the unlabeled samples and its corresponding cluster centers. The effectiveness of the proposed method is evaluated on two case studies, one is on an experimental bearing fault dataset from our laboratory test-rig, and the other is on a publicly dataset from a bearing degradation test. The comparison results on these two case studies demonstrate that the proposed method can perform better in bearing fault diagnosis under limited labeled samples than existing diagnostic methods.