Scholarlink

<p>Fracture traces are key to characterizing rock mass mechanical properties as their distribution and geometry directly affect the stability and safety of underground structures. Traditional detection methods struggle under deep underground conditions, such as uneven lighting, surface roughness variations, and complex textures. To address these challenges, this study integrates the SegFormer network featuring a hierarchical transformer structure and fracture feature extraction mechanisms to achieve intelligent detection and multi-dimensional characterization of deep hard-rock fracture traces. A comprehensive database containing 1616 hard-rock fracture images was constructed, with 80% used for model training. Experimental results demonstrate that the constructed SegFormer achieves superior performance (mIoU: 0.8193 ± 0.0704) in environmental noise suppression and boundary completeness, substantially outperforming U-Net (mIoU: 0.6781 ± 0.0780), DeepLabv3 + (mIoU: 0.7256 ± 0.0875), and traditional edge detection algorithms. Furthermore, a geometric simplification algorithm is applied to extract skeleton structures from segmented traces while preserving key topological features, enabling quantification of critical parameters including trace length, orientation angle, density, and intensity. Meanwhile, visualization analysis of attention mechanisms and feature maps enhanced model explainability, providing deep insights into the hierarchical feature extraction process. The effectiveness of the proposed method is validated through multiple case studies, successfully quantifying fracture traces under complex conditions and demonstrating its application potential for automatic detection and characterization of fracture traces in deep hard-rock environments.</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

Autonomous Fracture Trace Characterization in Underground Hard-Rock Environments via a Multi-scale Hierarchical Transformer Network

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Yingui Qiu,
  • Jian Zhou,
  • Chuanqi Li

摘要

Fracture traces are key to characterizing rock mass mechanical properties as their distribution and geometry directly affect the stability and safety of underground structures. Traditional detection methods struggle under deep underground conditions, such as uneven lighting, surface roughness variations, and complex textures. To address these challenges, this study integrates the SegFormer network featuring a hierarchical transformer structure and fracture feature extraction mechanisms to achieve intelligent detection and multi-dimensional characterization of deep hard-rock fracture traces. A comprehensive database containing 1616 hard-rock fracture images was constructed, with 80% used for model training. Experimental results demonstrate that the constructed SegFormer achieves superior performance (mIoU: 0.8193 ± 0.0704) in environmental noise suppression and boundary completeness, substantially outperforming U-Net (mIoU: 0.6781 ± 0.0780), DeepLabv3 + (mIoU: 0.7256 ± 0.0875), and traditional edge detection algorithms. Furthermore, a geometric simplification algorithm is applied to extract skeleton structures from segmented traces while preserving key topological features, enabling quantification of critical parameters including trace length, orientation angle, density, and intensity. Meanwhile, visualization analysis of attention mechanisms and feature maps enhanced model explainability, providing deep insights into the hierarchical feature extraction process. The effectiveness of the proposed method is validated through multiple case studies, successfully quantifying fracture traces under complex conditions and demonstrating its application potential for automatic detection and characterization of fracture traces in deep hard-rock environments.