Scholarlink

<p>The planetary transmission is an important component of vehicle power transmission, with complex and variable working conditions. High-accuracy edge intelligent online monitoring and fault diagnosis technologies are key to ensuring its reliable operation. This study proposes a dual label smoothing convolution network (DLSCN) to meet the requirement. First, using wavelet transform convolution (WTConv) to generate feature matrix, combined with cosine similarity between patches and attention mechanism between channels after cutting and encoding, an intricate teacher model with superior performance is constructed. Then, a light-weight student model is built with depthwise separable convolution (DWConv) and squeeze-and-excitation attention (SE attention), using label smoothing for first-order label smoothing. A response-based knowledge distillation framework is applied to transfer the smoothed label response from the teacher model to the student model for second-order label smoothing. An edge intelligent fault diagnosis system based on DLSCN (EC-DLSCN) is designed, which offloads the fault diagnosis task to the industrial three-proof edge device. The test and comparison results demonstrate that DLSCN achieves 12.6 % improvement in average diagnosis accuracy, 84.0 % improvement in inference speed, and 88.5 % improvement in average memory occupation, compared to current light-weight methods. Moreover, the online experiments and analysis results show that EC-DLSCN can effectively perform edge intelligent online fault diagnosis, with inference speed fully meeting the real-time requirements for 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

Implementation of edge intelligent fault diagnosis for planetary transmission based on dual label smoothing convolutional network

  • 期刊论文

  • 发表日期：

    2025年10月8日

查看全文
  • Jinxiao Li,
  • Ran Gong,
  • Yi Xu,
  • Wei Jiang

摘要

The planetary transmission is an important component of vehicle power transmission, with complex and variable working conditions. High-accuracy edge intelligent online monitoring and fault diagnosis technologies are key to ensuring its reliable operation. This study proposes a dual label smoothing convolution network (DLSCN) to meet the requirement. First, using wavelet transform convolution (WTConv) to generate feature matrix, combined with cosine similarity between patches and attention mechanism between channels after cutting and encoding, an intricate teacher model with superior performance is constructed. Then, a light-weight student model is built with depthwise separable convolution (DWConv) and squeeze-and-excitation attention (SE attention), using label smoothing for first-order label smoothing. A response-based knowledge distillation framework is applied to transfer the smoothed label response from the teacher model to the student model for second-order label smoothing. An edge intelligent fault diagnosis system based on DLSCN (EC-DLSCN) is designed, which offloads the fault diagnosis task to the industrial three-proof edge device. The test and comparison results demonstrate that DLSCN achieves 12.6 % improvement in average diagnosis accuracy, 84.0 % improvement in inference speed, and 88.5 % improvement in average memory occupation, compared to current light-weight methods. Moreover, the online experiments and analysis results show that EC-DLSCN can effectively perform edge intelligent online fault diagnosis, with inference speed fully meeting the real-time requirements for fault detection.