Scholarlink

<p>Rolling bearing fault diagnosis is critical for ensuring mechanical system reliability and safety. However, traditional diagnostic methods suffer from information loss due to the non-stationary and one-dimensional nature of vibration signals, which significantly reduces diagnostic accuracy. To address these challenges, this study proposes a time–frequency-domain fusion self-attention dual-stream network (TF-SADSN) for enhanced fault diagnosis. The network employs a dual-input architecture that synergistically integrates a Transformer module for capturing long-range temporal dependencies with an enhanced 2D CNN for extracting spatial and frequency-domain features. The dual-stream architecture leverages parallel computing to simultaneously process time-domain and frequency-domain information, improving computational efficiency when handling large-scale bearing datasets. Specifically, the Transformer utilizes self-attention mechanisms to identify sequential patterns in time-domain signals, while the enhanced 2D CNN incorporates multi-scale attention mechanisms to strengthen frequency-domain feature extraction from time–frequency spectrograms. Extensive experiments conducted on two benchmark bearing datasets demonstrate that TF-SADSN achieves superior fault classification accuracy compared to existing state-of-the-art methods, validating its effectiveness and potential for practical bearing fault diagnosis 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

Intelligent fault diagnosis for rolling bearings: a time–frequency-domain fusion self-attention dual-stream network approach

  • 期刊论文

  • 发表日期：

    2026年1月18日

查看全文
  • Dong Qiu,
  • Mimi Liu,
  • Zengpeng Lu,
  • Ziyang Zhang,
  • Changjun Guan,
  • Zesheng Lin,
  • Yan Li

摘要

Rolling bearing fault diagnosis is critical for ensuring mechanical system reliability and safety. However, traditional diagnostic methods suffer from information loss due to the non-stationary and one-dimensional nature of vibration signals, which significantly reduces diagnostic accuracy. To address these challenges, this study proposes a time–frequency-domain fusion self-attention dual-stream network (TF-SADSN) for enhanced fault diagnosis. The network employs a dual-input architecture that synergistically integrates a Transformer module for capturing long-range temporal dependencies with an enhanced 2D CNN for extracting spatial and frequency-domain features. The dual-stream architecture leverages parallel computing to simultaneously process time-domain and frequency-domain information, improving computational efficiency when handling large-scale bearing datasets. Specifically, the Transformer utilizes self-attention mechanisms to identify sequential patterns in time-domain signals, while the enhanced 2D CNN incorporates multi-scale attention mechanisms to strengthen frequency-domain feature extraction from time–frequency spectrograms. Extensive experiments conducted on two benchmark bearing datasets demonstrate that TF-SADSN achieves superior fault classification accuracy compared to existing state-of-the-art methods, validating its effectiveness and potential for practical bearing fault diagnosis applications.