Machine learning approach to railway ballast degradation prognosis considering crumb rubber modification and parent rock strength

Abstract

Parent rock strength and crumb rubber modification are two critical mechanical parameters that significantly decide the ballast layer degradation subjected to train dynamic loading. Using machine learning to predict ballast degradation considering these two parameters is helpful for deciding ballasted track maintenance cycle. In the current study, the ballast degradation process data (variables: parent rock types, loading types, ballast gradations and compositions of crumb rubber-ballast mixture) were used to train machine learning models. The drop-weight impact loading tests were performed to simulate different train dynamic loadings. Two well-established machine learning models, i.e., random forest (RF) and support vector regression (SVR) were trained and verified, to more effectively assess the importance of these variables. The results from the validated machine learning models confirm that the parent rock type is the most influential parameter, followed by the loading type (applied stress level), to control and predict the degradation of the ballast-CR mixture. The experimental assessment reveals that although the incorporation of CR suppresses degradation across all characterized rock types, the improvement in performance of the ballast-CR specimen against degradation is more noticeable for high-strength parent rock subjected to a considerable stress level. Meanwhile, this positive influence is also observed for ballast of weaker strength when the applied stress level is low.