Scholarlink

<p>Traditional methods for liquefaction assessment typically rely on critical curves, which are inadequate for effectively addressing the complex nonlinear relationships among factors influencing liquefaction. Meanwhile, existing machine learning models for liquefaction prediction often require extensive parameter tuning and feature engineering tailored to specific datasets to achieve satisfactory performance. Moreover, these models generally lack a detailed analysis of how liquefaction features influence prediction outcomes. This study develops an interpretable stacked model for predicting liquefaction potential using stacking algorithms and Standard Penetration Test (SPT) data. The eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Random Forest (RF) models were selected as base learners, while a Multilayer Perceptron (MLP) model was employed as the meta-learner. SHapley Additive exPlanations (SHAP) analysis and decision tree visualization were used to analyze the impact trends of input features on the final prediction results. To mitigate the impact of imbalanced data exacerbated by unreasonable sampling, the Borderline Synthetic Minority Over-sampling Technique (Borderline SMOTE) was introduced. Results show that the stacked model achieved an accuracy of 92.65%, significantly outperforming the pre-stacking models and the NCEER method. The interpretability analysis indicates that the influence trends of various factors in the proposed model align with existing knowledge, confirming the rationality and reliability of the model. Moreover, the corrected SPT blow count, fines content, cyclic shear stress ratio, and soil depth were identified as the four most important factors in machine learning-based liquefaction prediction models. When the corrected SPT blow count exceeds 25, the likelihood of soil liquefaction significantly decreases.</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

Interpretable liquefaction prediction model based on stacking algorithm

  • 期刊论文

  • 发表日期：

    2025年9月15日

查看全文
  • Xiao Long,
  • Rui Sun,
  • Tong Zheng,
  • Yu Wang,
  • Zhaocheng Zhong

摘要

Traditional methods for liquefaction assessment typically rely on critical curves, which are inadequate for effectively addressing the complex nonlinear relationships among factors influencing liquefaction. Meanwhile, existing machine learning models for liquefaction prediction often require extensive parameter tuning and feature engineering tailored to specific datasets to achieve satisfactory performance. Moreover, these models generally lack a detailed analysis of how liquefaction features influence prediction outcomes. This study develops an interpretable stacked model for predicting liquefaction potential using stacking algorithms and Standard Penetration Test (SPT) data. The eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Random Forest (RF) models were selected as base learners, while a Multilayer Perceptron (MLP) model was employed as the meta-learner. SHapley Additive exPlanations (SHAP) analysis and decision tree visualization were used to analyze the impact trends of input features on the final prediction results. To mitigate the impact of imbalanced data exacerbated by unreasonable sampling, the Borderline Synthetic Minority Over-sampling Technique (Borderline SMOTE) was introduced. Results show that the stacked model achieved an accuracy of 92.65%, significantly outperforming the pre-stacking models and the NCEER method. The interpretability analysis indicates that the influence trends of various factors in the proposed model align with existing knowledge, confirming the rationality and reliability of the model. Moreover, the corrected SPT blow count, fines content, cyclic shear stress ratio, and soil depth were identified as the four most important factors in machine learning-based liquefaction prediction models. When the corrected SPT blow count exceeds 25, the likelihood of soil liquefaction significantly decreases.