Prediction of Slope Failure through Integrating Statistical Design of Experiments and Artificial Neural Networks

This study proposes the integration of analytical geotechnical methods and statistical tools to develop a prediction model for the factor of safety (FS) of homogenous soil slopes as a function of soil and geometry properties. The proposed model is developed by combining statistical design of experiment (DOE), artificial neural networks (ANN), and limit equilibrium (LE) analysis to generate suitable combinations of input factors and for data analysis. The model adequacy as a prediction tool for preliminary design is evaluated by measuring accuracy using a case histories dataset. The performance results indicate that the predicted values of FS have a high correlation with the computer-simulated values (analytical values), indicating that the developed model compares to the use of performing analysis in LE software without the need for special packages. However, all the compared tools fall into the low accuracy zone if the threshold between stability and failure is set to 1. For achieving high accuracy (area under the curve >0.85) from the proposed classifier, a safety margin of 20% should be used. In other words, for a 1.20 FS threshold between stability and failure, the proposed model classifies with high accuracy for the analyzed cases.

Keywords: Algorithms, Data processing, Factor of safety, Slope stability, Artificial neural networks, Statistical design of experiment (DOE).