Scholarlink

<p>Under complex working conditions, traditional fault detection methods have limitations like many parameters and complex calculations. To solve this, a bearing fault detection model based on smooth dilated convolution and shuffling algorithm was proposed. It uses smooth convolution kernels to capture local vibration-signal features, reduces computational complexity via group convolution and channel washing, simplifies the structure with network pruning and knowledge distillation, and combines bidirectional gated recurrent units and generative adversarial networks to capture long-term dependencies. Compared with existing methods, it significantly cuts the number of model parameters and reasoning time while keeping detection accuracy. Experimental data shows that in the sample classification task, its accuracy rate is 97.88%, average reasoning time is 274 fps, computational cost is 1.66 FLOPs, and parameter quantity is 7.76&#xa0;M, all better than comparison models. In bearing feature extraction and fault detection tasks, its average fitting accuracy is 96.13% and detection accuracy is 99.62%, also better than comparison models. The research suggests the model can balance model lightweighting and detection performance, and is suitable for real-time fault monitoring in resource-constrained 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

Bearing fault detection with lightweight feature extraction mechanism based on smoothed dilated convolution

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月15日

查看全文
  • Yufeng Pang,
  • Xiaojuan Li

摘要

Under complex working conditions, traditional fault detection methods have limitations like many parameters and complex calculations. To solve this, a bearing fault detection model based on smooth dilated convolution and shuffling algorithm was proposed. It uses smooth convolution kernels to capture local vibration-signal features, reduces computational complexity via group convolution and channel washing, simplifies the structure with network pruning and knowledge distillation, and combines bidirectional gated recurrent units and generative adversarial networks to capture long-term dependencies. Compared with existing methods, it significantly cuts the number of model parameters and reasoning time while keeping detection accuracy. Experimental data shows that in the sample classification task, its accuracy rate is 97.88%, average reasoning time is 274 fps, computational cost is 1.66 FLOPs, and parameter quantity is 7.76 M, all better than comparison models. In bearing feature extraction and fault detection tasks, its average fitting accuracy is 96.13% and detection accuracy is 99.62%, also better than comparison models. The research suggests the model can balance model lightweighting and detection performance, and is suitable for real-time fault monitoring in resource-constrained scenarios.