LAFD-Net: Learning With Noisy Pseudo-Labels for Semisupervised Bearing Fault Diagnosis

Fault diagnosis for the rolling bearing is an important field that has received increasing attention in recent years. The main challenge for this task is the lack of labeled data. Existing works circumvent this problem with pseudo-labels generated from labeled data. However, these pseudo-labels are...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE sensors journal 2023-02, Vol.23 (4), p.3911-3923
Hauptverfasser: Jian, Yifan, Chen, Zhi, Lei, Yinjie, He, Zhengxi, Zhao, Yang, He, Liang, Luo, Wei, Chen, Xuekun
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Fault diagnosis for the rolling bearing is an important field that has received increasing attention in recent years. The main challenge for this task is the lack of labeled data. Existing works circumvent this problem with pseudo-labels generated from labeled data. However, these pseudo-labels are noisy even with consistency checks or confidence-based filtering due to the minimal amount of training data. To solve this problem, a novel label-level antinoise fault diagnosis network (LAFD-Net) is proposed in this article. Specifically, we propose an online asymptotic label updating (OALU) strategy that contains two updating stages: self-correction stage and cross-correction stage. The proposed OALU can stably and reliably generate new corrected pseudo-labels, gradually replacing the old noisy ones. The LAFD-Net adopts a student–teacher architecture. For such a student–teacher model, we propose a consistency enhancement (CE) loss to strengthen the feature consistency between the student and teacher networks, aiming to achieve more efficient use of plentiful unlabeled data via feature regularization. Finally, a series of experiments were conducted using the University of Cincinnati Intelligent Maintenance System (IMS) Center dataset and the Case Western Reserve University (CWRU) bearing dataset. The experimental results demonstrate that the proposed semisupervised learning (SSL) schemes outperformed existing state-of-the-art methods with the same percentage of labeled data samples.
ISSN:1530-437X
1558-1748
DOI:10.1109/JSEN.2023.3233957