Scholarlink

<p>Accurate identification of the evolutionary stages of tunnel water and mud inrush disasters is a critical challenge for ensuring construction safety. Existing methods often suffer from reliance on single-threshold criteria, sparse monitoring data, and difficulty in label acquisition. To address these issues, this study proposes an intelligent prediction framework that integrates multi-source time-series data with semi-supervised transfer learning. Four key monitoring parameters—face water inflow (<i>Q</i>), crown displacement (<i>D</i>), pore water pressure (<i>u</i>), and surrounding rock stress (<i>σ</i>)—are collected to construct a multi-dimensional dynamic feature set. A hybrid 1D convolutional neural network&#xa0;(1D-CNN) and gated recurrent unit (GRU) architecture is adopted to extract deep temporal features. A source–target domain transfer mechanism is then introduced, where simulation data serve as the source domain and in situ monitoring data as the target domain. Distribution alignment is achieved using maximum mean discrepancy. To address label scarcity in the target domain, a confidence-guided pseudo-labeling strategy is employed to augment the fine-tuning dataset. Experimental results demonstrate that the proposed method achieves a perfect Macro F1-score of 1.0 on the three-stage classification task, outperforming support vector machine with radial basis function kernel and long short-term memory with bidirectional GRU models by 43.0% and 9.5%, respectively. This study significantly enhances the accuracy and robustness of disaster stage recognition and offers a viable solution and theoretical support for intelligent early warning under complex geological conditions.</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

Intelligent Identification of Evolutionary Stages of Tunnel Water and Mud Inrush Hazards Based on Multi-source Time-Series Data Fusion: A Transfer Learning and Semi-supervised Learning Approach

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Xiayi Yao,
  • Liucheng Yu,
  • Yinchuan Bai,
  • Fashun shi,
  • Dong Xu,
  • Mingli Huang

摘要

Accurate identification of the evolutionary stages of tunnel water and mud inrush disasters is a critical challenge for ensuring construction safety. Existing methods often suffer from reliance on single-threshold criteria, sparse monitoring data, and difficulty in label acquisition. To address these issues, this study proposes an intelligent prediction framework that integrates multi-source time-series data with semi-supervised transfer learning. Four key monitoring parameters—face water inflow (Q), crown displacement (D), pore water pressure (u), and surrounding rock stress (σ)—are collected to construct a multi-dimensional dynamic feature set. A hybrid 1D convolutional neural network (1D-CNN) and gated recurrent unit (GRU) architecture is adopted to extract deep temporal features. A source–target domain transfer mechanism is then introduced, where simulation data serve as the source domain and in situ monitoring data as the target domain. Distribution alignment is achieved using maximum mean discrepancy. To address label scarcity in the target domain, a confidence-guided pseudo-labeling strategy is employed to augment the fine-tuning dataset. Experimental results demonstrate that the proposed method achieves a perfect Macro F1-score of 1.0 on the three-stage classification task, outperforming support vector machine with radial basis function kernel and long short-term memory with bidirectional GRU models by 43.0% and 9.5%, respectively. This study significantly enhances the accuracy and robustness of disaster stage recognition and offers a viable solution and theoretical support for intelligent early warning under complex geological conditions.