Scholarlink

Purpose <p>This study proposes a novel ResNeXt50-based fault diagnosis method for gearbox, which integrates Depthwise Separable Convolution (DSC) and Convolutional Block Attention Module (CBAM) to mitigate the challenges of high computational complexity and low diagnostic efficiency in existing approaches.</p> Methods <p>The proposed methodology initially employs Continuous Wavelet Transform (CWT) to transform vibration signals into two-dimensional time–frequency representations, thereby enhancing feature discriminability. The approach further incorporates Depthwise Separable Convolution (DSC) in the second layer of ResNeXt50's residual blocks to optimize computational efficiency, while positioning a Convolutional Block Attention Module (CBAM) before the pooling layer. This CBAM integration strengthens critical feature extraction through dynamic channel-wise and spatial weight adaptation.</p> Results <p>The proposed model achieves rolling bearing classification accuracies of 96.125% and 95.375%, and gear classification accuracies of 95.625% and 94.625%, showing a significant improvement over the conventional ResNeXt50 baseline and other comparative models. The model's superiority is further corroborated by confusion matrix analysis and T-SNE visualization, with all evaluation metrics (including accuracy, recall, and F1-score) consistently exceeding 91%. These comprehensive results substantiate the method's practical utility and diagnostic reliability for rolling bearing and gear fault detection.</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

Deep Learning-Based Fault Diagnosis of Gearbox Using CBAM-DSC-ResNeXt50 Structure

  • 期刊论文

  • 发表日期：

    2025年7月10日

查看全文
  • Shuihai Dou,
  • Fangyi Dai,
  • Yanping Du,
  • Fu Liu,
  • Ting Li,
  • Dechen Yao,
  • Huijuan Bai

摘要

Purpose

This study proposes a novel ResNeXt50-based fault diagnosis method for gearbox, which integrates Depthwise Separable Convolution (DSC) and Convolutional Block Attention Module (CBAM) to mitigate the challenges of high computational complexity and low diagnostic efficiency in existing approaches.

Methods

The proposed methodology initially employs Continuous Wavelet Transform (CWT) to transform vibration signals into two-dimensional time–frequency representations, thereby enhancing feature discriminability. The approach further incorporates Depthwise Separable Convolution (DSC) in the second layer of ResNeXt50's residual blocks to optimize computational efficiency, while positioning a Convolutional Block Attention Module (CBAM) before the pooling layer. This CBAM integration strengthens critical feature extraction through dynamic channel-wise and spatial weight adaptation.

Results

The proposed model achieves rolling bearing classification accuracies of 96.125% and 95.375%, and gear classification accuracies of 95.625% and 94.625%, showing a significant improvement over the conventional ResNeXt50 baseline and other comparative models. The model's superiority is further corroborated by confusion matrix analysis and T-SNE visualization, with all evaluation metrics (including accuracy, recall, and F1-score) consistently exceeding 91%. These comprehensive results substantiate the method's practical utility and diagnostic reliability for rolling bearing and gear fault detection.