Scholarlink

<p>Leakage is a common issue in tunnel structures, and its investigation and detection are crucial for ensuring the safety of the tunnel structure. A consumer-grade multi-sensor, i.e. the Azure Kinect, is introduced herein to collect a multi-sensor dataset, including RGB and IR images under normal lighting, as well as IR images under low lighting conditions. The advanced directional connectivity-based segmentation network (DconnNet) is then implemented for leakage detection. Quantitative and qualitative comparisons with traditional U-Net and Attention U-Net (AU-Net) models are both carried out. The results show that the detection performance of the infrared dataset outperforms the RGB dataset, with IR images under normal lighting conditions achieving the best performance. DconnNet achieved an mIoU value and F1 Score of 85.98% and 92.43%, respectively. Compared to U-Net and AU-Net, the mIoU value and F1 Score increased by 10.9% and 10.4%, respectively. The proposed method in this study is applicable for detecting multiple targets and various shapes of leakage areas. The combination of low-cost sensors and advanced detection networks can significantly reduce the costs of the hardware and network development optimization, and meanwhile achieve accurate leakage 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

A Consumer-Grade Multi-sensor Deep Learning Approach for Tunnel Leakage Detection

  • 期刊论文

  • 发表日期：

    2025年7月5日

查看全文
  • Yuanzhen Xu,
  • Shuqi Ma,
  • Xiangchen Yao,
  • Zhaoyuan Zhang,
  • Qilin Yan

摘要

Leakage is a common issue in tunnel structures, and its investigation and detection are crucial for ensuring the safety of the tunnel structure. A consumer-grade multi-sensor, i.e. the Azure Kinect, is introduced herein to collect a multi-sensor dataset, including RGB and IR images under normal lighting, as well as IR images under low lighting conditions. The advanced directional connectivity-based segmentation network (DconnNet) is then implemented for leakage detection. Quantitative and qualitative comparisons with traditional U-Net and Attention U-Net (AU-Net) models are both carried out. The results show that the detection performance of the infrared dataset outperforms the RGB dataset, with IR images under normal lighting conditions achieving the best performance. DconnNet achieved an mIoU value and F1 Score of 85.98% and 92.43%, respectively. Compared to U-Net and AU-Net, the mIoU value and F1 Score increased by 10.9% and 10.4%, respectively. The proposed method in this study is applicable for detecting multiple targets and various shapes of leakage areas. The combination of low-cost sensors and advanced detection networks can significantly reduce the costs of the hardware and network development optimization, and meanwhile achieve accurate leakage detection.