Scholarlink

<p>This study presents an automated decision-support framework for pavement maintenance and rehabilitation (M&amp;R) prioritization using machine learning and explainable artificial intelligence (AI). A stacked ensemble classifier—combining tuned Light gradient boosting machine (LightGBM), ExtraTrees, support vector machine (SVM), and multilayer perceptron (MLP) as base learners with XGBoost as the meta-learner—was developed using nine pavement condition indicators with a dataset of 1166 points. The model predicts discrete M&amp;R classes with strong predictive performance on the studied dataset, achieving an accuracy of 98.29%, macro F1-score of 0.98, precision of 0.963, and recall of 0.964. Model transparency was achieved through SHAP (SHapley Additive exPlanations), which identified Current million standard axles (MSA), Roughness International Roughness Index (IRI), and Structural Number as dominant predictors. A SHAP-informed composite scoring system enabled intra-class prioritization of road segments into Urgent, Medium, and Low categories. Sensitivity analysis confirmed that rankings for urgent interventions remained stable under ± 20% feature contributions variations. The framework offers a robust and interpretable framework for network-level pavement management, supporting data-driven intervention planning with improved transparency and objectivity.</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

Stacked Ensemble Learning with SHAP-Prioritized Intervention Ranking for Pavement Maintenance Automation

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Aakash Gupta,
  • Pradeep Kumar

摘要

This study presents an automated decision-support framework for pavement maintenance and rehabilitation (M&R) prioritization using machine learning and explainable artificial intelligence (AI). A stacked ensemble classifier—combining tuned Light gradient boosting machine (LightGBM), ExtraTrees, support vector machine (SVM), and multilayer perceptron (MLP) as base learners with XGBoost as the meta-learner—was developed using nine pavement condition indicators with a dataset of 1166 points. The model predicts discrete M&R classes with strong predictive performance on the studied dataset, achieving an accuracy of 98.29%, macro F1-score of 0.98, precision of 0.963, and recall of 0.964. Model transparency was achieved through SHAP (SHapley Additive exPlanations), which identified Current million standard axles (MSA), Roughness International Roughness Index (IRI), and Structural Number as dominant predictors. A SHAP-informed composite scoring system enabled intra-class prioritization of road segments into Urgent, Medium, and Low categories. Sensitivity analysis confirmed that rankings for urgent interventions remained stable under ± 20% feature contributions variations. The framework offers a robust and interpretable framework for network-level pavement management, supporting data-driven intervention planning with improved transparency and objectivity.