Prediction of rail ballast breakage using a hybrid ML methodology

Abstract

Particle breakage is a key performance indicator to estimate ballast degradation as it severely affects the performance and maintenance of rail tracks. Most constitutive models usually based on continuum mechanics have rarely been able to estimate the rate and intensity of particle degradation under repeated wheel loading. In this regard, this paper presents a novel model for predicting ballast breakage under prolonged cyclic loading using artificial neural networks (ANN) coupled with a genetic algorithm (GA), hence the acronym GA-ANN. For this study, a comprehensive database consisting of 130 experimental datasets on ballast breakage under cyclic loading conditions is used. Unlike most black-box type machine learning (ML) models, this study incorporates a knowledge-guided selection of 9 input parameters encompassing gradation characteristics, particle angularity, the initial physical state of the granular assembly, and the applied stress state. To overcome limitations associated with potential overfitting when using smaller datasets of the Ballast Breakage Index (BBI), this study employs an innovative approach by integrating k-fold cross-validation and regularization with conventional GA-ANN algorithm. The proposed GA-ANN model showed superior performance in predicting BBI at different loading cycles and proved to be 50% more efficient when compared to conventional ANN and other ML techniques. When verified against unseen laboratory and field data, the GA-ANN model yielded an $R^2$ between 0.85 and 0.95, thus proving its broader capability. Further, global sensitivity analysis is performed to identify the most significant parameters (cyclic deviatoric stress, number of load cycles and frequency) which warrant more attention during maintenance.