Scholarlink

<p>Semi-supervised learning has emerged as a promising paradigm for rotating machinery fault diagnosis with limited annotations. However, real world industrial scenarios are inevitably plagued by severe class imbalance and the emergence of out-of-distribution (OOD) faults, which significantly undermine the reliability of standard SSL methods. Existing approaches often implicitly assume balanced distributions or closed-set conditions, leading to biased predictions and confident errors on unknown faults. To bridge this gap, this paper proposes SemiViT-CL, a unified semi-supervised framework that integrates a Vision Transformer (ViT) backbone with contrastive learning to ensure robust diagnosis under such complex conditions. Specifically, we introduce a class-progress-aware pseudo labeling strategy that dynamically calibrates confidence thresholds to mitigate the bias toward majority classes. Furthermore, to prevent the model from overfitting to OOD samples, we design a consistency-guided OOD filtering mechanism, which leverages the discrepancy between weak and strong augmentations to detect and reject unknown faults effectively. Extensive experiments on the SEU and PU benchmarks demonstrate that SemiViT-CL outperforms state-of-the-art methods, achieving an average accuracy of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(&gt; 98\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>&gt;</mo> <mn>98</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, along with superior macro-averaged F1-scores and balanced accuracy, even under low label ratios. The results validate the method’s capability to maintain high diagnostic precision and OOD detection robustness, suggesting its potential for reliable condition monitoring in real world industrial environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Semi-supervised Vision Transformer with Contrastive Learning and OOD Filtering for Fault Diagnosis of Rotating Machinery

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Lin Zhu,
  • Zhangjun Wu,
  • Shiyuan Zheng,
  • Haifeng Ling

摘要

Semi-supervised learning has emerged as a promising paradigm for rotating machinery fault diagnosis with limited annotations. However, real world industrial scenarios are inevitably plagued by severe class imbalance and the emergence of out-of-distribution (OOD) faults, which significantly undermine the reliability of standard SSL methods. Existing approaches often implicitly assume balanced distributions or closed-set conditions, leading to biased predictions and confident errors on unknown faults. To bridge this gap, this paper proposes SemiViT-CL, a unified semi-supervised framework that integrates a Vision Transformer (ViT) backbone with contrastive learning to ensure robust diagnosis under such complex conditions. Specifically, we introduce a class-progress-aware pseudo labeling strategy that dynamically calibrates confidence thresholds to mitigate the bias toward majority classes. Furthermore, to prevent the model from overfitting to OOD samples, we design a consistency-guided OOD filtering mechanism, which leverages the discrepancy between weak and strong augmentations to detect and reject unknown faults effectively. Extensive experiments on the SEU and PU benchmarks demonstrate that SemiViT-CL outperforms state-of-the-art methods, achieving an average accuracy of \(> 98\%\) > 98 % , along with superior macro-averaged F1-scores and balanced accuracy, even under low label ratios. The results validate the method’s capability to maintain high diagnostic precision and OOD detection robustness, suggesting its potential for reliable condition monitoring in real world industrial environments.