Scholarlink

<p>The stability of surrounding rock is a critical challenge in underground engineering. Accurate prediction of Excavation Damaged Zone Thickness (EDZT) in surrounding rock is vital for designing safe underground support systems. Conventional prediction methods frequently encounter limitations in terms of accuracy and generalisation, attributable to the intricate geology and variable coupling effects characteristic of the region. The study proposes a novel prediction model by integrating an improved Grey Wolf Optimizer (IGWO) with a Multiple-Kernel Relevance Vector Machine (MKRVM). Unlike standard algorithms, the IGWO avoids local optima, while the MKRVM enhances modeling flexibility. The model was developed using 209 field datasets with four inputs: Rock Mass Strength (RMS), Joint Index (JI), Embedding Depth (ED), and Drift Span (DS). Experimental results demonstrate that the IGWO-MKRVM achieves superior performance, with coefficients of determination (R²) of 0.9552 and 0.9528, and Root Mean Square Errors (RMSE) of 0.1356 and 0.0932 for training and testing sets, respectively. It outperforms Back Propagation Neural Network (BPNN), Artificial Neural Network (ANN), Gaussian Process Regression (GPR), Sparrow Search Algorithm-Support Vector Machine (SSA-SVM), and Dragonfly Algorithm-Random Forest (DA-RF) in terms of accuracy and generalisation. Finally, Shapley Additive Explanations (SHAP) analysis reveals that JI and ED are the most influential variables affecting EDZT.</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

An interpretable IGWO-MKRVM model for predicting excavation damaged zone thickness of drift

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月14日

查看全文
  • Ruzi Yang,
  • Guangquan Zhang,
  • Yicheng Ye,
  • Mingchao Wan

摘要

The stability of surrounding rock is a critical challenge in underground engineering. Accurate prediction of Excavation Damaged Zone Thickness (EDZT) in surrounding rock is vital for designing safe underground support systems. Conventional prediction methods frequently encounter limitations in terms of accuracy and generalisation, attributable to the intricate geology and variable coupling effects characteristic of the region. The study proposes a novel prediction model by integrating an improved Grey Wolf Optimizer (IGWO) with a Multiple-Kernel Relevance Vector Machine (MKRVM). Unlike standard algorithms, the IGWO avoids local optima, while the MKRVM enhances modeling flexibility. The model was developed using 209 field datasets with four inputs: Rock Mass Strength (RMS), Joint Index (JI), Embedding Depth (ED), and Drift Span (DS). Experimental results demonstrate that the IGWO-MKRVM achieves superior performance, with coefficients of determination (R²) of 0.9552 and 0.9528, and Root Mean Square Errors (RMSE) of 0.1356 and 0.0932 for training and testing sets, respectively. It outperforms Back Propagation Neural Network (BPNN), Artificial Neural Network (ANN), Gaussian Process Regression (GPR), Sparrow Search Algorithm-Support Vector Machine (SSA-SVM), and Dragonfly Algorithm-Random Forest (DA-RF) in terms of accuracy and generalisation. Finally, Shapley Additive Explanations (SHAP) analysis reveals that JI and ED are the most influential variables affecting EDZT.