Joint learning system based on semi–pseudo–label reliability assessment for weak–fault diagnosis with few labels

• Published in: Mechanical systems and signal processing Vol. 189; p. 110089
• Main Authors: Gao, Da-wei, Zhu, Yong-sheng, Yan, Ke, Fu, Hong, Ren, Zhi-jun, Kang, Wei, Guedes Soares, C.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.04.2023
• Subjects: Label noise; Pseudo–label; Sample reliability; Semi–supervised learning; Weak–fault diagnosis; Sample reliability; Pseudo–label; Label noise; Weak–fault diagnosis; Semi–supervised learning
• ISSN: 0888-3270; 1096-1216
• DOI: 10.1016/j.ymssp.2022.110089

•Deep learning method for extracting weak–fault–related features when the labels are insufficient.•Pseudo–label selection mechanism based on reliability assessment.•Joint learning workflow combining the advantages of UL, TL, and SSL.•SSL based on multiple views considering prior knowledge of signal processing. Deep neural networks exhibit excellent performance in fault feature extraction for considerable amounts of data. However, data labeling is a difficult task in practical engineering, which may lead to problems in fault diagnosis particularly when faults are weak. To resolve the foregoing, a semi–pseudo–labeling diagnosis system is proposed in this paper. The proposed system considers the confidence and reliability of samples to cope with situations where labels are insufficient and faults are weak. By adding pseudo–labels, unlabeled data whose fault information is swamped by a large amount of noise can achieve low–density separation and entropy regularization in the sample space. Consequently, the training of deep learning models for weak–fault diagnosis is supported. Regarding the traditional pseudo–labeling problems in weak–fault–related feature extraction, a series of solutions has been proposed to solve the problems in the field of fault diagnosis. The designed model reduces pseudo–label noise and enhances the capability of weak–fault–related feature extraction. The effectiveness of this method was validated on the datasets collected by simulating faulty bearings and those sustaining actual failure.