Scholarlink

<p>In full-face tunnel boring machine construction, reliable lithology identification at the tunnel face is essential for optimizing boring parameters and reducing cutter wear. To address the challenge of insufficient labeled data in early-stage projects, this study employs a lithology recognition approach based on self-supervised learning. The Simple framework for Contrastive Learning of Visual Representations (SimCLR) model is employed to extract lithology-relevant features without labels, followed by performance evaluation through downstream classification tasks. A rock muck image acquisition system was established at the Yinchao-Jiliao 2–3 tunnel project, resulting in a dataset comprising 18,149 images across six lithological classes. This study focuses on analyzing key visual factors influencing model performance and proposes targeted optimization strategies. The results demonstrate that: (1) Extracting single rock fragments from muck images for lithology recognition should be prioritized, as it ensures equivalent accuracy while simultaneously reducing redundant information and computational costs. (2) Color information significantly improves recognition accuracy, with color images achieving up to a 20% improvement in accuracy compared to grayscale images. (3) Incorporating cross-project auxiliary data for pretraining enables the use of unlabeled public datasets, enhancing recognition performance in engineering with limited labeled data. Based on these findings, it is recommended that practical lithology recognition models focus on single fragments features, preserve image color, and utilize cross-domain data for pretraining to improve adaptability in lithology recognition.</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

Analysis and Optimization of Key Factors in Self-Supervised Lithology Recognition from TBM Muck Images

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Zi-Kai Dong,
  • Zhong-Sheng Tan,
  • Xu Li,
  • Zhen-Liang Zhou,
  • Guo-Shuai Tian

摘要

In full-face tunnel boring machine construction, reliable lithology identification at the tunnel face is essential for optimizing boring parameters and reducing cutter wear. To address the challenge of insufficient labeled data in early-stage projects, this study employs a lithology recognition approach based on self-supervised learning. The Simple framework for Contrastive Learning of Visual Representations (SimCLR) model is employed to extract lithology-relevant features without labels, followed by performance evaluation through downstream classification tasks. A rock muck image acquisition system was established at the Yinchao-Jiliao 2–3 tunnel project, resulting in a dataset comprising 18,149 images across six lithological classes. This study focuses on analyzing key visual factors influencing model performance and proposes targeted optimization strategies. The results demonstrate that: (1) Extracting single rock fragments from muck images for lithology recognition should be prioritized, as it ensures equivalent accuracy while simultaneously reducing redundant information and computational costs. (2) Color information significantly improves recognition accuracy, with color images achieving up to a 20% improvement in accuracy compared to grayscale images. (3) Incorporating cross-project auxiliary data for pretraining enables the use of unlabeled public datasets, enhancing recognition performance in engineering with limited labeled data. Based on these findings, it is recommended that practical lithology recognition models focus on single fragments features, preserve image color, and utilize cross-domain data for pretraining to improve adaptability in lithology recognition.