Scholarlink

<p>Given that rolling bearing vibration signals exhibit non-linearity and non-stationarity, coupled with strong noise, this renders the extraction of bearing fault features using conventional vibration analysis methods exceedingly challenging. To address these issues, this paper proposes an intelligent fault diagnosis model based on stacking variable step-size multiscale fuzzy dispersion entropy (SVSMFDE) and MDSCNN-Transformer. Firstly, the proposed SVSMFDE employs a stacking variable step-size strategy to minimise information loss and enhance feature extraction capability. For the MDSCNN-Transformer, we introduce an enhanced CNN-Transformer approach. This method employs MDSCNN to extract sparse masks, guiding sparse attention mechanisms to effectively fuse local and global information. This resolves the Transformer’s limitations in local feature extraction and the CNN’s difficulties in modelling global dependencies, ultimately improving classification stability and accuracy. The proposed model framework’s effectiveness is validated across two datasets. Experimental results demonstrate classification accuracies of 99.59 % and 99.25 %, respectively, significantly outperforming other benchmark models.</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 rolling bearing fault diagnosis method based on stacking variable step-size multiscale fuzzy dispersion entropy and MDSCNN-Transformer

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Ze Ouyang,
  • Qingkang Wang,
  • Shuai Zhou,
  • Ruqi Song,
  • Yichao Yin,
  • Haoxuan Zhu,
  • Zhuo Chen

摘要

Given that rolling bearing vibration signals exhibit non-linearity and non-stationarity, coupled with strong noise, this renders the extraction of bearing fault features using conventional vibration analysis methods exceedingly challenging. To address these issues, this paper proposes an intelligent fault diagnosis model based on stacking variable step-size multiscale fuzzy dispersion entropy (SVSMFDE) and MDSCNN-Transformer. Firstly, the proposed SVSMFDE employs a stacking variable step-size strategy to minimise information loss and enhance feature extraction capability. For the MDSCNN-Transformer, we introduce an enhanced CNN-Transformer approach. This method employs MDSCNN to extract sparse masks, guiding sparse attention mechanisms to effectively fuse local and global information. This resolves the Transformer’s limitations in local feature extraction and the CNN’s difficulties in modelling global dependencies, ultimately improving classification stability and accuracy. The proposed model framework’s effectiveness is validated across two datasets. Experimental results demonstrate classification accuracies of 99.59 % and 99.25 %, respectively, significantly outperforming other benchmark models.