Repose angle prediction of railway ballast based on Hopper Flow Test and PCA-Stacking ensemble learning method

Abstract

The deterioration of track ballast is mainly reflected in the fragmentation of ballast particles and the evolution of grading. The repose angle of the ballast is the internal friction angle, reflecting the shear strength of the track ballast. The angle of repose of 158 ballast samples with varying gradation parameters was tested by conducting hopper flow tests on ballast particles. Then, based on this database, a stacking ensemble regressor-based principal component analysis (PCA) was proposed for the repose angle prediction. The data are first processed by PCA and then randomized into the train (80%) and test (20%) data. Besides stacking ensemble regressor, five different individual regressors, including support vector regressor, k-nearest neighbors, random forest, gradient boosting decision tree, and adaptive boosting, are compared and analyzed. The results indicate that the stacking ensemble regressor performs better than single regressors and demonstrates enhanced robustness and generalization capabilities. Additionally, the Effect of the number of principal component indicators on the prediction accuracy is determined, and the difference in stacking ensemble prediction performance with and without PCA is discussed. Finally, a computer prediction system for ballast repose angle was developed to make the proposed stacking ensemble regressor directly applicable to the railroad construction site. This model can be used in practical track maintenance decisions, whereby the ballast gradation can facilitate rapid assessment of the angle of repose.