<p>The scarcity of labeled fault samples in real operating environments severely limits the deployment of bearing fault diagnosis models trained on laboratory data. Although Maximum Classifier Discrepancy (MCD)–based unsupervised domain adaptation can reduce cross-domain feature mismatch and form well-structured target clusters, it implicitly assumes consistent semantic correspondence between classifier outputs and physical fault categories. In practice, this assumption is often violated: target samples may exhibit high permutation-invariant clustering accuracy (PICA) but suffer from systematic semantic permutation, leading to a large semantic misalignment gap (SMG). To address this issue, we propose a Semantic Calibration–enhanced Maximum Classifier Discrepancy framework (SC-MCD) for cross-domain bearing fault diagnosis with few-shot target supervision. A multi-scale CNN–LSTM encoder extracts discriminative temporal–spectral representations, and MCD performs unsupervised feature adaptation. A lightweight K-shot semantic calibration module then estimates an optimal label permutation via Hungarian matching to realign output semantics while keeping the learned features frozen. Experiments on laboratory-to-real bearing transfer tasks show that SC-MCD effectively resolves semantic inconsistency. With only one labeled target sample per class, SC-MCD aligns accuracy with PICA and drives SMG close to zero; moreover, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(K\!\approx \!3\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>K</mi> <mspace width="-0.166667em" /> <mo>≈</mo> <mspace width="-0.166667em" /> <mn>3</mn> </mrow> </math></EquationSource> </InlineEquation> provides strong robustness to moderate anchor noise.</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

SC-MCD: semantic calibration with maximum classifier discrepancy for cross-domain bearing fault diagnosis using few-shot target samples

  • Zhenfei Li,
  • Tongwen Yuan,
  • Xin Liu,
  • Kai Gong,
  • Ling Zhang

摘要

The scarcity of labeled fault samples in real operating environments severely limits the deployment of bearing fault diagnosis models trained on laboratory data. Although Maximum Classifier Discrepancy (MCD)–based unsupervised domain adaptation can reduce cross-domain feature mismatch and form well-structured target clusters, it implicitly assumes consistent semantic correspondence between classifier outputs and physical fault categories. In practice, this assumption is often violated: target samples may exhibit high permutation-invariant clustering accuracy (PICA) but suffer from systematic semantic permutation, leading to a large semantic misalignment gap (SMG). To address this issue, we propose a Semantic Calibration–enhanced Maximum Classifier Discrepancy framework (SC-MCD) for cross-domain bearing fault diagnosis with few-shot target supervision. A multi-scale CNN–LSTM encoder extracts discriminative temporal–spectral representations, and MCD performs unsupervised feature adaptation. A lightweight K-shot semantic calibration module then estimates an optimal label permutation via Hungarian matching to realign output semantics while keeping the learned features frozen. Experiments on laboratory-to-real bearing transfer tasks show that SC-MCD effectively resolves semantic inconsistency. With only one labeled target sample per class, SC-MCD aligns accuracy with PICA and drives SMG close to zero; moreover, \(K\!\approx \!3\) K 3 provides strong robustness to moderate anchor noise.