A consistency regularization based semi-supervised learning approach for intelligent fault diagnosis of rolling bearing. (1st December 2020)

• Record Type: Journal Article
• Authors: Yu, Kun; Ma, Hui; Lin, Tian Ran; Li, Xiang
• Is Part Of: Measurement. Volume 165(2020)
• Publication Date: 2020-12-01
• Subjects: Deep learning -- Semi-supervised learning -- Consistency regularization -- Fault diagnosis; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2020.107987
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A SSL method for intelligent fault diagnosis of rolling bearing is proposed. Data augmentation is the basic foundation of the proposed method. Two consistency terms make the trained network less sensitive to the extra perturbations. The proposed method achieves excellent performance on the experimental study. Abstract: Deep learning has been widely used nowadays to achieve an automated fault diagnosis of rolling bearings. However, most of deep learning based bearing fault diagnosis methods are based on the assumption that the recorded samples are labeled data, though most of field data are recorded without label information. To address this issue, an effective semi-supervised learning method based on the principle of consistency regularization is proposed in this study. The principle of consistency regularization underlines that the model predictions should be less sensitive to the extra perturbation imposed on the input samples. In the proposed method, a data augmentation method is proposed to serve as the extra perturbation, which is imposed on both the labeled and unlabeled samples to enrich the data library. Meanwhile, a label predicting process is formulated to estimate the appreciable label distribution for the unlabeled sample. Correspondingly, two consistency loss terms are introduced to regularize the model predictions for both the labeled and unlabeled samples to be invariant to the extra perturbation, among which a supervised loss term is adopted to enforce the model predictions for augmented labeled samples to be consistent with its true label information and an unsupervised loss is proposed to minimize the discrepancy between label distributions for the original unlabeled samples and its appreciable label distributions. The analysis result on an experimental bearing fault dataset demonstrates that the proposed method can provide an excellent identification performance under limited labeled samples situation.