Scholarlink

In mines, coal must be transported over long distances via conveyor belts to the surface. However, foreign objects such as gravel chunks and anchors within the fast-moving coal stream can damage or tear the belt, and may even obstruct the coal discharge opening, critically impairing the safety and efficiency of mine transport. To address the current challenges of low recognition accuracy and subpar real-time performance in coal mine foreign object detection, this paper introduces a novel classification approach. This method integrates image preprocessing based on Canny edge detection with an optimized Swin-Transformer model. The preprocessing includes three stages: adaptive luminance adjustment, Canny edge detection, and adaptive image fusion, aimed at accentuating crucial edge features to bolster the model’s geometric structure recognition capabilities. Additionally, we embed a lightweight Circular Grouped Attention (CGA) module into the Swin-Transformer, merging channel and spatial attentions while balancing dense and sparse attentions to augment the model’s integrative information processing. Testing on the CUMT-Belt dataset, our approach reaches a classification accuracy of 96.9%, evidencing its potential to significantly enhance coal mine transport safety and efficiency.

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

Foreign Object Classification for Coal Conveyor Belts Based on Deep Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Siyu Chen,
  • Mingtao Pei

摘要

In mines, coal must be transported over long distances via conveyor belts to the surface. However, foreign objects such as gravel chunks and anchors within the fast-moving coal stream can damage or tear the belt, and may even obstruct the coal discharge opening, critically impairing the safety and efficiency of mine transport. To address the current challenges of low recognition accuracy and subpar real-time performance in coal mine foreign object detection, this paper introduces a novel classification approach. This method integrates image preprocessing based on Canny edge detection with an optimized Swin-Transformer model. The preprocessing includes three stages: adaptive luminance adjustment, Canny edge detection, and adaptive image fusion, aimed at accentuating crucial edge features to bolster the model’s geometric structure recognition capabilities. Additionally, we embed a lightweight Circular Grouped Attention (CGA) module into the Swin-Transformer, merging channel and spatial attentions while balancing dense and sparse attentions to augment the model’s integrative information processing. Testing on the CUMT-Belt dataset, our approach reaches a classification accuracy of 96.9%, evidencing its potential to significantly enhance coal mine transport safety and efficiency.