Scholarlink

<p>Micro-particle contamination is a key issue in automotive manufacturing, as it directly affects system reliability and cleanliness. This study proposes a BP–CNN framework for particle classification across different size ranges. Particles larger than 300&#xa0;μm are classified using a back-propagation (BP) neural network based on geometric descriptors, while particles in the range of 50–300&#xa0;μm are classified using a DenseNet121-based convolutional neural network (CNN) that learns morphological features from images. The BP-CNN framework classified 7,500 particle images from 13 classes, collected from various automotive transmission manufacturing processes such as CNC machining, casting, high-pressure water jet deburring, brush deburring, and robot reaming … Experimental results show that the proposed BP–CNN framework achieves higher performance than standalone BP, CNN ResNet, CNN EfficientNet, CNN-TL-SVM model, with an overall accuracy of 99.12%. This research provides performance address for automotive particle identification.</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

Micro-particle classification using deep learning back propagation and cnn algorithm

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Trong-Dat Huynh,
  • Minh-Thuan Tran,
  • Huy-Tuan Pham,
  • Quang-Khoa Dang,
  • Thien-Ngon Dang,
  • Quoc-Bao Phan

摘要

Micro-particle contamination is a key issue in automotive manufacturing, as it directly affects system reliability and cleanliness. This study proposes a BP–CNN framework for particle classification across different size ranges. Particles larger than 300 μm are classified using a back-propagation (BP) neural network based on geometric descriptors, while particles in the range of 50–300 μm are classified using a DenseNet121-based convolutional neural network (CNN) that learns morphological features from images. The BP-CNN framework classified 7,500 particle images from 13 classes, collected from various automotive transmission manufacturing processes such as CNC machining, casting, high-pressure water jet deburring, brush deburring, and robot reaming … Experimental results show that the proposed BP–CNN framework achieves higher performance than standalone BP, CNN ResNet, CNN EfficientNet, CNN-TL-SVM model, with an overall accuracy of 99.12%. This research provides performance address for automotive particle identification.