Abstract
The assessment of loess slope stability is a highly complex nonlinear problem. There are many factors that influence the stability of loess slopes. Some of them have the characteristic of uncertainty. Meanwhile, the relationship between different factors may be complicated. The existence of multiple correlation will affect the objectivity of stability analysis and prevent the model from making correct judgments. In this paper, the main factors affecting the stability of loess slopes are analyzed by means of the partial least-squares regression (PLSR). After that, two new synthesis variables with better interpretation to the dependent variables are extracted. By this way, the multicollinearity among variables is overcome preferably. Moreover, the BP neural network is further used to determine the nonlinear relationship between the new components and the slope safety factor. Then, a new improved BP model based on the partial least-squares regression, which is initialized by the particle swarm optimization (PSO) algorithm, is developed, i.e., the PLSR-BP model. The network with global convergence capability is simplified and more efficient. The test results of the model show satisfactory precision, which indicates that the model is feasible and effective for stability evaluation of loess slopes.