Scholarlink

<p>Rolling bearings are vital components in mechanical systems, and early fault diagnosis is crucial for equipment safety and reliability. Traditional empirical and statistical methods show limited accuracy for non-stationary vibration signals and poor robustness under noise or small-sample conditions. To address this, a fault diagnosis method based on Transformer energy entropy features and MShOA-SVM is proposed. The Transformer performs end-to-end additive decomposition to extract multi-scale features, and the energy entropy of intrinsic mode components (IMFs) forms a high-dimensional feature vector. The mantis shrimp optimization algorithm optimizes SVM parameters for improved classification. Experiments on the CWRU dataset achieve 99.31 % accuracy, 98.56 % specificity, and an AUC of 0.991. Ablation results confirm that Transformer-based feature extraction and MShOA significantly enhance performance. The method maintains strong robustness under noise and small-sample conditions, demonstrating effective fault identification and practical value for online monitoring and intelligent maintenance.</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

Rolling bearing fault diagnosis method based on transformer energy entropy and MShOA-SVM

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Jiajun Zhang,
  • Jie Chen,
  • Fen Chen,
  • Ming Ma

摘要

Rolling bearings are vital components in mechanical systems, and early fault diagnosis is crucial for equipment safety and reliability. Traditional empirical and statistical methods show limited accuracy for non-stationary vibration signals and poor robustness under noise or small-sample conditions. To address this, a fault diagnosis method based on Transformer energy entropy features and MShOA-SVM is proposed. The Transformer performs end-to-end additive decomposition to extract multi-scale features, and the energy entropy of intrinsic mode components (IMFs) forms a high-dimensional feature vector. The mantis shrimp optimization algorithm optimizes SVM parameters for improved classification. Experiments on the CWRU dataset achieve 99.31 % accuracy, 98.56 % specificity, and an AUC of 0.991. Ablation results confirm that Transformer-based feature extraction and MShOA significantly enhance performance. The method maintains strong robustness under noise and small-sample conditions, demonstrating effective fault identification and practical value for online monitoring and intelligent maintenance.