Scholarlink

<p>In the industrial environment, strong noise interference and insufficient training samples bring severe challenges to mechanical fault diagnosis. A fault diagnosis framework based on a novel multimodal data fusion method and an improved multiscale residual network is proposed to overcome the above issues. Firstly, a multimodal data fusion method using principal component analysis and feature map concatenation is introduced. It significantly enhances the feature retention of different modal signals and improves their adaptability to various rotating machinery types. Secondly, by integrating a shifted window attention mechanism, the proposed multiscale residual network demonstrates an improved feature extraction and global information perception, exhibiting a superior fault classification capability than comparison models. Finally, three cases are utilized to test the applicability of the framework for bearings and motors, showcasing its robustness under high noise and limited training sample scenarios. Besides, the framework shows a generalization capability with over 99% accuracy in multiple working conditions, reflecting its potential for industrial applications.</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

SWR2 Net: a fault diagnosis framework for rotating machine under limited samples and noise interference

  • 期刊论文

  • 发表日期：

    2025年5月24日

查看全文
  • Yuan Zhuang,
  • Deqiang He,
  • Zhenzhen Jin,
  • Zhexian Wei,
  • Haimeng Sun,
  • Jinxing Wu,
  • Wang Zhuang

摘要

In the industrial environment, strong noise interference and insufficient training samples bring severe challenges to mechanical fault diagnosis. A fault diagnosis framework based on a novel multimodal data fusion method and an improved multiscale residual network is proposed to overcome the above issues. Firstly, a multimodal data fusion method using principal component analysis and feature map concatenation is introduced. It significantly enhances the feature retention of different modal signals and improves their adaptability to various rotating machinery types. Secondly, by integrating a shifted window attention mechanism, the proposed multiscale residual network demonstrates an improved feature extraction and global information perception, exhibiting a superior fault classification capability than comparison models. Finally, three cases are utilized to test the applicability of the framework for bearings and motors, showcasing its robustness under high noise and limited training sample scenarios. Besides, the framework shows a generalization capability with over 99% accuracy in multiple working conditions, reflecting its potential for industrial applications.