Scholarlink

<p>This study compared six backpropagation-based artificial neural networks (ANN) and two kernel-based least squares support vector machine (LSSVM) models to obtain an optimal performance model to estimate the ultimate bearing capacity (UBC) of shallow foundations. A database was collected from the literature and utilized to train (78 datasets) and test (19 datasets) each model. For the first time, the effect of multicollinearity of internal friction angle (ɸ), soil unit weight (ɣ), length to width ratio of footing (L/B), footing depth (D), and footing width (B) was analyzed by the variance inflation factor (VIF) method in estimating UBC. The study revealed that the radial basis function (RBF)-based LSSVM achieved an accuracy of over 99% with the least residuals, i.e., root mean square error of 67.2377&#xa0;kPa, and outperformed the linear (L) kernel-based LSSVM and Levenberg–Marquardt (LM)-based ANN models. A novel relationship plot, i.e., a feature multicollinearity-sensitivity relationship plot, revealed that the RBF_LSSVM model experienced overfitting (= 1.42) due to the weak multicollinearity of B (= 2.38) and D (= 2.02), which contributed 25.612% (for D) and 21.284% (for B), respectively.</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

Estimation of Bearing Capacity of Shallow Foundations on Cohesionless Soil Using Backpropagation Neural Networks and LSSVM Models

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • Jitendra Khatti,
  • Kamaldeep Singh Grover,
  • Hyeong-Joo Kim,
  • Kevin Bagas A. Mawuntu,
  • Tae-Woong Park

摘要

This study compared six backpropagation-based artificial neural networks (ANN) and two kernel-based least squares support vector machine (LSSVM) models to obtain an optimal performance model to estimate the ultimate bearing capacity (UBC) of shallow foundations. A database was collected from the literature and utilized to train (78 datasets) and test (19 datasets) each model. For the first time, the effect of multicollinearity of internal friction angle (ɸ), soil unit weight (ɣ), length to width ratio of footing (L/B), footing depth (D), and footing width (B) was analyzed by the variance inflation factor (VIF) method in estimating UBC. The study revealed that the radial basis function (RBF)-based LSSVM achieved an accuracy of over 99% with the least residuals, i.e., root mean square error of 67.2377 kPa, and outperformed the linear (L) kernel-based LSSVM and Levenberg–Marquardt (LM)-based ANN models. A novel relationship plot, i.e., a feature multicollinearity-sensitivity relationship plot, revealed that the RBF_LSSVM model experienced overfitting (= 1.42) due to the weak multicollinearity of B (= 2.38) and D (= 2.02), which contributed 25.612% (for D) and 21.284% (for B), respectively.