Semi-supervised gearbox fault diagnosis under variable working conditions based on masked contrastive learning
To address the problem that it is difficult to label variable working condition gearbox fault samples and the significant data distribution discrepancies in practical engineering, which result in reduced accuracy of fault diagnosis models, a semi-supervised gearbox fault diagnosis method based on ma...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Editorial Office of Journal of Mechanical Strength
2025-06-01
|
| Series: | Jixie qiangdu |
| Subjects: | |
| Online Access: | http://www.jxqd.net.cn/thesisDetails#DOI：10.16579/j.issn.1001.9669.2025.06.009 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | To address the problem that it is difficult to label variable working condition gearbox fault samples and the significant data distribution discrepancies in practical engineering, which result in reduced accuracy of fault diagnosis models, a semi-supervised gearbox fault diagnosis method based on masked contrastive learning is proposed. Firstly, a random mask was used to hide part of the information in the unlabeled dataset, generating two different masked instances for each unlabeled sample. Secondly, a dynamic convolutional neural network was employed to dynamically weight and aggregate the masked instances, enabling discriminative feature modeling of different masked instances. Then, a contrastive learning framework was constructed with the optimization goal of maximizing the similarity between features of different masked instances. By enhancing the consistency of feature representations of masked instance pairs, the model's dependency on labels was reduced. Finally, during the fine-tuning phase, a domain-conditioned feature correction strategy was introduced to generate target domain feature corrections. By aligning source domain features and target domain corrected features according to the metric of minimizing domain feature distribution discrepancies, the method explicitly reduces the domain distribution differences caused by varying working conditions. Validation on a variable working condition gearbox fault dataset demonstrates the effectiveness of the proposed method. |
|---|---|
| ISSN: | 1001-9669 |