Scholarlink

<p>Small sample sizes and robust noise continue to limit the accuracy of deep learning (DL)–based bearing fault diagnosis (FD). To overcome these challenges, we propose two novel modules: Cross-layer Calibration Inception (CLC-Inception)—which combines multiscale 1D convolutions with a Squeeze-and-Excitation attention block—and Channel-Spatial Dynamic Recalibration Block (CSDRB)—comprising a Channel Recalibration Block (CRB) and a Spatial Recalibration Block (SRB) that adaptively reweights spatial and temporal features. In our method, raw vibration signals first pass through a wide convolutional layer for enhanced noise reduction; next, the Cross-layer Calibration Inception extracts deep multiscale features, which are then modified using residual convolutional layers; finally, the CSDRB prioritizes critical diagnostic information. The effectiveness of the proposed method is validated on two publicly available bearing datasets: the Case Western Reserve University (CWRU) dataset and the Lanzhou University of Technology (LUT) dataset. Comparative analyses and ablation experiments confirm that the proposed approach achieves 99.34% accuracy under small sample conditions and an average of 96.16% accuracy under strong noise, surpassing four advanced deep learning methods. These results highlight the robustness and potential applicability of the developed method for real-world bearing fault diagnosis scenarios.</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

Efficient fault diagnosis method based on dynamic recalibration mechanism with improved inception network for rolling bearings under limited samples

  • 期刊论文

  • 发表日期：

    2025年9月2日

查看全文
  • Alaeldden Abduelhadi,
  • Jie Cao,
  • Haopeng Liang,
  • Shanqin Yuan

摘要

Small sample sizes and robust noise continue to limit the accuracy of deep learning (DL)–based bearing fault diagnosis (FD). To overcome these challenges, we propose two novel modules: Cross-layer Calibration Inception (CLC-Inception)—which combines multiscale 1D convolutions with a Squeeze-and-Excitation attention block—and Channel-Spatial Dynamic Recalibration Block (CSDRB)—comprising a Channel Recalibration Block (CRB) and a Spatial Recalibration Block (SRB) that adaptively reweights spatial and temporal features. In our method, raw vibration signals first pass through a wide convolutional layer for enhanced noise reduction; next, the Cross-layer Calibration Inception extracts deep multiscale features, which are then modified using residual convolutional layers; finally, the CSDRB prioritizes critical diagnostic information. The effectiveness of the proposed method is validated on two publicly available bearing datasets: the Case Western Reserve University (CWRU) dataset and the Lanzhou University of Technology (LUT) dataset. Comparative analyses and ablation experiments confirm that the proposed approach achieves 99.34% accuracy under small sample conditions and an average of 96.16% accuracy under strong noise, surpassing four advanced deep learning methods. These results highlight the robustness and potential applicability of the developed method for real-world bearing fault diagnosis scenarios.