Abstract: The collapsibility of the loess is an important engineering characteristic, and the mechanism of collapsibility is very complicated. With the increasing of engineering construction in the Western Inner Mongolia, more and more attention has been paid to the collapsible deformation characteristics of the loess. This paper presents a case history of the loess study in the Baotou area, a typical loess area in Western Mongolia. The statistical analysis on more than 200 loess samples from ten projects are performed and the influencing factors of the loess collapsibility are evaluated by using the mathematical statistics. The natural water content, dry weight, void ratio, saturation, plasticity index and compression modulus are determined as the multiple factors in the loess collapsibility prediction. On this basis, the principal component analysis (PCA) is used to eliminate the overlap among the influencing factors, so that the recombined factors are independent of each other. Then, the artificial neural network (ANN) is used to train the prediction model. Finally, the accuracy of the prediction model is verified by the actual project data. The relative errors of the measured and the predicted collapsibility are -12.5% and -10.4% respectively. The collapsibility grades are therefore consistent. The results show that the method for the loess collapsibility analysis can be used as an important reference in the engineering practice.

Key words: Loess, Collapsible Deformation, Collapsible Coefficient, Neural Network, Principal Component Analysis