Scholarlink

<p>Accurately delineating rock mass excavatability is a prerequisite for optimizing geotechnical engineering construction strategies. The classic excavatability chart, while offering intuitive engineering guidance, is founded on empirical statistical adjustments that remain vulnerable to subjective bias. More importantly, the limited interpretability inherent in traditional methodologies poses challenges in rapidly updating the graphic’s classification scheme, underscoring the need for advanced predictive techniques. In contrast to classical statistical fitting approaches, soft computing techniques, unconstrained by rigid priori assumptions, excel at uncovering nonlinear patterns in coupled geological datasets. Exploiting this advantage, this study systematically evaluates the feasibility and reliability of machine learning techniques for rock mass excavatability prediction based on an expanded database comprising 214 samples. Ten classifiers are rigorously evaluated to avoid narrow perspectives in model selection, among which the multilayer perceptron (MLP) achieved superior performance, and its accuracy is further enhanced to 88.89% through global hyperparameter optimization via the genetic algorithm (GA‑MLP). To enhance decision transparency, SHapley Additive exPlanations (SHAP) were employed to reveal the nonlinear coupling between rock strength and structural spacing that governs excavatability decisions. After data‑driven remapping of the feature plane with GA‑MLP, the misclassification rate drops from 12 to 8%, and statistical analysis confirmed that the updated boundaries better align with the sample distribution. In addition, gene expression programming (GEP) generated explicit probabilistic discriminants, recasting the target problem into a soft classification framework to resolve threshold ambiguities in hard classification. By introducing a fuzzy margin parameter, a more conservative five-level soft classification scheme was proposed that provides engineers with enhanced decision flexibility. Finally, the modeling outcomes were encapsulated in a graphical user interface that facilitates automatic excavatability grading, thereby furnishing practitioners with an intuitive and interpretable tool for continuously updating the excavatability chart in response to newly acquired site data.</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

Cutting—Edge Soft Computing Technologies for Rock Mass Excavatability: Transforming Prediction with Hybrid GA—MLP and GEP—Based Criteria

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Shuai Huang,
  • Jian Zhou

摘要

Accurately delineating rock mass excavatability is a prerequisite for optimizing geotechnical engineering construction strategies. The classic excavatability chart, while offering intuitive engineering guidance, is founded on empirical statistical adjustments that remain vulnerable to subjective bias. More importantly, the limited interpretability inherent in traditional methodologies poses challenges in rapidly updating the graphic’s classification scheme, underscoring the need for advanced predictive techniques. In contrast to classical statistical fitting approaches, soft computing techniques, unconstrained by rigid priori assumptions, excel at uncovering nonlinear patterns in coupled geological datasets. Exploiting this advantage, this study systematically evaluates the feasibility and reliability of machine learning techniques for rock mass excavatability prediction based on an expanded database comprising 214 samples. Ten classifiers are rigorously evaluated to avoid narrow perspectives in model selection, among which the multilayer perceptron (MLP) achieved superior performance, and its accuracy is further enhanced to 88.89% through global hyperparameter optimization via the genetic algorithm (GA‑MLP). To enhance decision transparency, SHapley Additive exPlanations (SHAP) were employed to reveal the nonlinear coupling between rock strength and structural spacing that governs excavatability decisions. After data‑driven remapping of the feature plane with GA‑MLP, the misclassification rate drops from 12 to 8%, and statistical analysis confirmed that the updated boundaries better align with the sample distribution. In addition, gene expression programming (GEP) generated explicit probabilistic discriminants, recasting the target problem into a soft classification framework to resolve threshold ambiguities in hard classification. By introducing a fuzzy margin parameter, a more conservative five-level soft classification scheme was proposed that provides engineers with enhanced decision flexibility. Finally, the modeling outcomes were encapsulated in a graphical user interface that facilitates automatic excavatability grading, thereby furnishing practitioners with an intuitive and interpretable tool for continuously updating the excavatability chart in response to newly acquired site data.