Scholarlink

<p>Fault diagnosis of rolling bearings is crucial for industrial equipment operation, with the core challenge lying in extracting features that can effectively distinguish between different fault types. Existing methods ignore the manifold structure of the signal and lack the ability to preserve geometric consistency, which causes the distortion of structure during the feature extraction process, thereby impairing the model’s diagnostic accuracy. This study introduces a novel approach for rolling bearing fault diagnosis, employing a deep learning network that incorporates manifold structure preservation to address this issue. Time–frequency features are first extracted using continuous wavelet transform, and symmetric positive definite matrices are constructed to combine the local and global features of the signals. Subsequently, a deep learning network with manifold structure preserved, containing nonlinear encoding and dimensionality reduction, is established to extract the robustness features. In addition, a joint loss function incorporating manifold structure preservation and discriminative learning is designed to further enhance the model’s classification precision. The diagnostic performance of the proposed model is validated through two experiments, achieving a classification accuracy of 99%, outperforming other baseline methods. It also exhibits superior robustness under high-noise conditions, confirming its reliability in challenging scenarios. These results underscore its potential for practical application in real-world mechanical fault diagnosis.</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 structure-preserving deep manifold network for rolling bearing fault diagnosis

  • 期刊论文

  • 发表日期：

    2025年8月2日

查看全文
  • Hongdi Zhou,
  • Jun Luo,
  • Yijie Cai

摘要

Fault diagnosis of rolling bearings is crucial for industrial equipment operation, with the core challenge lying in extracting features that can effectively distinguish between different fault types. Existing methods ignore the manifold structure of the signal and lack the ability to preserve geometric consistency, which causes the distortion of structure during the feature extraction process, thereby impairing the model’s diagnostic accuracy. This study introduces a novel approach for rolling bearing fault diagnosis, employing a deep learning network that incorporates manifold structure preservation to address this issue. Time–frequency features are first extracted using continuous wavelet transform, and symmetric positive definite matrices are constructed to combine the local and global features of the signals. Subsequently, a deep learning network with manifold structure preserved, containing nonlinear encoding and dimensionality reduction, is established to extract the robustness features. In addition, a joint loss function incorporating manifold structure preservation and discriminative learning is designed to further enhance the model’s classification precision. The diagnostic performance of the proposed model is validated through two experiments, achieving a classification accuracy of 99%, outperforming other baseline methods. It also exhibits superior robustness under high-noise conditions, confirming its reliability in challenging scenarios. These results underscore its potential for practical application in real-world mechanical fault diagnosis.