Prediction of resilient modulus of ballast under cyclic loading using machine learning techniques

Abstract

The resilient modulus (MR) of ballast is one of the key output parameters in any rail design project because it controls the elastic magnitude of track deformation under cyclic loading. This study investigates the response of MR under cyclic conditions as a function of four key parameters, i.e., the loading magnitude, the number of loading cycles, the loading frequency, and the confining pressure. To do so, two non-linear predictive models, namely, the artificial neural network (ANN), and the adaptive neuro-fuzzy inference system (ANFIS), are used to predict the MR values under different loading conditions. To evaluate and predict MR, an experimental database with 196 data samples is considered in this study. A series of sensitivity analyses is carried out to investigate the most effective parameters in each non-linear model and also predict the highest performance model. Although the results from the primary validation phase are satisfactory for the ANN and ANFIS models, ANFIS proves better (i.e., the coefficient of determination = 0.709) at estimating the MR during the secondary validation phase, using an independent dataset. Hence, it can be used as a powerful and practical model for predicting the magnitude of MR. On the basis of the ANFIS model, this study also offers some design considerations in terms of MR of ballast under a practical range of cyclic loading parameters.