Preliminary Discussion Regarding SVM Kernel Function Selection in the Twofold Rock Slope Prediction Model

Abstract

The kernel function, which is an important component of support vector machine (SVM) theory, directly affects the results of a prediction model. When establishing an effective prediction slope model, analysis factors such as slope angle, slope height, potential sliding body height and inclination, and cohesion and friction angle of each potential sliding surface need to be considered. As the results of an example design show, there is an appropriate regularity between analysis factors and kernel functions. For example, the radial basis function (RBF) kernel function is suitable for the geometry factors of a rock slope analysis, whereas the Sigmoid kernel function is better than RBF for analyzing the cohesion and friction angle of the back potential sliding surface; likewise, the linear kernel function is suitable for the material factors of a bottom sliding surface analysis. For these reasons, a combination of kernel functions is necessary for an overall analysis of complex rock slope problems. A comprehensive kernel function based on the analysis of different factors is proposed in this paper. Notably, the maximum absolute error of the test results using this comprehensive kernel function is only 0.1698, meaning that a comprehensive kernel function better embodies the failure mechanism of the rock slope when building a support vector machine (SVM) prediction model. Furthermore, the application results for the right bank slope of Dagang Mountain show that the comprehensive kernel function can reflect actual instability.