Scholarlink

<p>Accurate estimation of the rate of penetration (ROP) is essential for planning and optimizing tunnel boring machine (TBM) operations. Unlike previous studies that evaluated each project independently, this study develops a unified cross-project prediction framework using geotechnical datasets from the Queens Water Tunnel (United States, US) and the Karaj–Tehran Water Conveyance Tunnel (Iran). Three datasets—US only, Iran only, and a combined dataset—are constructed to examine how geological variability and data diversity influence prediction accuracy, model generalizability, and overfitting behavior. Twelve machine-learning (ML) algorithms are benchmarked under consistent preprocessing, fixed hyperparameters, and a unified tenfold cross-validation scheme. The results show that artificial neural networks (ANN) achieve the highest overall predictive accuracy, while tree-based ensemble models tend to overfit single-project datasets. All empirical and ML models exhibit reduced transferability when trained on a single project, confirming the strong dependence of ROP prediction on site-specific geological conditions. Integrating the two datasets improves the robustness and cross-project performance of most ML models, although it increases prediction variance due to enhanced heterogeneity. The findings highlight the critical role of dataset composition in ML-based TBM performance prediction and provide practical guidance for selecting reliable ROP estimation methods under varying rock 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

Comparison of performance prediction of TBM in two rock conditions using different machine learning techniques

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月28日

查看全文
  • Rui Yong,
  • Sujith Mangalathu,
  • Pengpeng Ni,
  • Changshuo Wang

摘要

Accurate estimation of the rate of penetration (ROP) is essential for planning and optimizing tunnel boring machine (TBM) operations. Unlike previous studies that evaluated each project independently, this study develops a unified cross-project prediction framework using geotechnical datasets from the Queens Water Tunnel (United States, US) and the Karaj–Tehran Water Conveyance Tunnel (Iran). Three datasets—US only, Iran only, and a combined dataset—are constructed to examine how geological variability and data diversity influence prediction accuracy, model generalizability, and overfitting behavior. Twelve machine-learning (ML) algorithms are benchmarked under consistent preprocessing, fixed hyperparameters, and a unified tenfold cross-validation scheme. The results show that artificial neural networks (ANN) achieve the highest overall predictive accuracy, while tree-based ensemble models tend to overfit single-project datasets. All empirical and ML models exhibit reduced transferability when trained on a single project, confirming the strong dependence of ROP prediction on site-specific geological conditions. Integrating the two datasets improves the robustness and cross-project performance of most ML models, although it increases prediction variance due to enhanced heterogeneity. The findings highlight the critical role of dataset composition in ML-based TBM performance prediction and provide practical guidance for selecting reliable ROP estimation methods under varying rock conditions.