Scholarlink

<p>Existing deep learning-based unsupervised domain adaptation (UDA) studies mostly measure only distribution discrepancies of independent feature layers, and the multiple layers of nesting of deep networks can lead to insufficient mining ability of relevant features. A novel UDA transfer diagnosis method with joint central moment discrepancy (JCMD) is proposed. Firstly, an effective channel attention mechanism (ECAM) is embedded into a convolutional neural network (CNN) to extract features related to bearing defect states. Secondly, a loss function based on JCMD is designed to measure joint distribution discrepancies between domains by multiplying central moment discrepancies of hidden activation values across multiple feature layers in both source and target domains. Finally, the effectiveness of the proposed method is validated on two rolling bearing datasets. The findings indicate that the proposed method effectively extracts fault features across various working conditions, enhancing transfer diagnosis performance in unsupervised cross-domain scenarios.</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

A novel unsupervised domain adaptation method with joint central moment discrepancy for fault diagnosis of rolling bearings

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Ran Ren,
  • Tao Liu,
  • Zhenya Wang,
  • Jiabing Gu

摘要

Existing deep learning-based unsupervised domain adaptation (UDA) studies mostly measure only distribution discrepancies of independent feature layers, and the multiple layers of nesting of deep networks can lead to insufficient mining ability of relevant features. A novel UDA transfer diagnosis method with joint central moment discrepancy (JCMD) is proposed. Firstly, an effective channel attention mechanism (ECAM) is embedded into a convolutional neural network (CNN) to extract features related to bearing defect states. Secondly, a loss function based on JCMD is designed to measure joint distribution discrepancies between domains by multiplying central moment discrepancies of hidden activation values across multiple feature layers in both source and target domains. Finally, the effectiveness of the proposed method is validated on two rolling bearing datasets. The findings indicate that the proposed method effectively extracts fault features across various working conditions, enhancing transfer diagnosis performance in unsupervised cross-domain scenarios.