Application of LS-SVM Method in Probabilistic Stability Analysis of Saturated Soil Slopes-Paper

Paper

Pdf Summary

The study evaluates least-squares support vector machines (LS-SVM) for estimating the factor of safety (FS) and probability of failure (PF) of a multilayer saturated soil slope. Because conventional limit equilibrium analyses require many repeated calculations for probabilistic assessment, the authors combine LS-SVM with Latin hypercube sampling (LHS) as a computationally efficient surrogate.<br /><br />The model uses soil unit weight, cohesion, and friction angle as inputs and FS as its output. It is tested on a multilayer slope, with results compared against direct Monte Carlo simulations using the Spencer and simplified Bishop methods. With 200 LS-SVM training samples, predicted failure probabilities were close to those from 30,000 direct simulations: 15.38% versus 15.95% for Spencer, and 30.73% versus 31.21% for Bishop. This suggests substantial potential savings in computational effort while retaining useful accuracy.<br /><br />Training-data selection is critical. Uniform sampling of cohesion and friction angle produced more reliable predictions across failure conditions than normal distributions, which concentrated samples near average values and performed poorly for low FS. The study also finds that increasing the training set can improve performance up to a point.<br /><br />The method has important limitations: predictions may be inaccurate when failure mechanisms are insufficiently represented in the training data. In particular, the first, sandy layer was treated deterministically, so potential failures associated with its uncertainty could not be learned. The authors recommend representing the probabilistic properties of every slope layer and using field and laboratory data to develop and validate training sets. Overall, LS-SVM appears promising for probabilistic slope stability analysis, provided its training data adequately capture soil variability and possible failure modes.

Keywords

least-squares support vector machines

LS-SVM

factor of safety

probability of failure

multilayer saturated soil slope

Latin hypercube sampling

Monte Carlo simulation

slope stability analysis

Spencer method

simplified Bishop method