Effect of SVM Kernel Functions on Bearing Capacity Assessment of Deep Foundations

Abstract

Pile foundations are vastly utilized in construction projects where their capacities (pile bearing capacity, PBC) should be determined in different stages of construction. A highly reliable and accurate prediction model can lead to many advantages, such as reducing the construction cost, shortening the construction timeline, and providing safety construction. Hence, the aim of this study is the developments of statistical and artificial intelligence (AI) models for predicting bearing capacities of 141 piles. At the preliminary of the study, features or inputs of this study to predict PBC were selected trough simple regression analysis. Then, this study presents different kernels of support vector machine (SVM) technique, i.e., the dot, the radial basis function (RBF), the polynomial, the neural, and the ANOVA to predict the PBC. The aforementioned models were evaluated by several performance indices and their results were compared using a simple ranking system. The results showed that the SVM-RBF model is able to achieve the highest coefficient of determination, R2 values which are 0.967 and 0.993 for training and testing stages, respectively. It is important to mention that a multiple regression model was also employed to predict PBC values. The other SVM kernels were provided a high degree of accuracy for estimating PBC, however, the SVM-RBF model is recommended to be used as a powerful, highly reliable, and simple solution for PBC prediction.

Keywords

• deep foundation capacity
• support vector machine
• kernels
• linear and multiple regression