Scholarlink

<p>Accurate prediction of soil unit weight is crucial for geotechnical engineering and soil characterization. This study leverages five advanced machine learning algorithms—Multi-Layer Perceptron (MLP), Random Forest (RF), Support Vector Regression (SVR), XGBoost, and AdaBoost with RF as a weak learner—to predict soil unit weight. Hyperparameters are optimized using randomized search cross-validation (RSCV), and model performance is evaluated using mean absolute error (MAE), root mean square error (RMSE), and R² metrics. The input features include soil depth (D), moisture content (MC), fine content (FC), cone tip resistance (QC), and cone local resistance (FS). An autoencoder-based feature augmentation technique is applied to enhance model performance. Before augmentation, AdaBoost with RF achieves the best performance (R² = 0.896), while SVR performs the worst (R² = 0.740). Post-augmentation, all models improve, with AdaBoost showing the highest R² and SVR achieving significant gains (R² = 0.782). SHAP analysis identifies D as the most critical feature, while QC and FS negatively impact accuracy. The results highlight AdaBoost with RF as the most effective algorithm for predicting soil unit weight, underscoring the value of feature augmentation in capturing complex patterns.</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

Use of CPT and other parameters for estimating soil unit weight using optimised machine learning models

  • 期刊论文

  • 发表日期：

    2025年1月31日

查看全文
  • Swaranjit Roy,
  • Abrar Rahman Abir,
  • Mehedi A. Ansary

摘要

Accurate prediction of soil unit weight is crucial for geotechnical engineering and soil characterization. This study leverages five advanced machine learning algorithms—Multi-Layer Perceptron (MLP), Random Forest (RF), Support Vector Regression (SVR), XGBoost, and AdaBoost with RF as a weak learner—to predict soil unit weight. Hyperparameters are optimized using randomized search cross-validation (RSCV), and model performance is evaluated using mean absolute error (MAE), root mean square error (RMSE), and R² metrics. The input features include soil depth (D), moisture content (MC), fine content (FC), cone tip resistance (QC), and cone local resistance (FS). An autoencoder-based feature augmentation technique is applied to enhance model performance. Before augmentation, AdaBoost with RF achieves the best performance (R² = 0.896), while SVR performs the worst (R² = 0.740). Post-augmentation, all models improve, with AdaBoost showing the highest R² and SVR achieving significant gains (R² = 0.782). SHAP analysis identifies D as the most critical feature, while QC and FS negatively impact accuracy. The results highlight AdaBoost with RF as the most effective algorithm for predicting soil unit weight, underscoring the value of feature augmentation in capturing complex patterns.