Perpustakaan Digital ITB

Track irregularities are the main disturbance to rail vehicle dynamics systems. Severe track irregularities will subject the passing vehicles to safety risks and unfavourable riding condition. Thus, track irregularities must be assessed regularly and kept within an acceptable level. Currently, track irregularities are commonly assessed based on measurement data obtained from special track measurement vehicles. However, track irregularity assessment methods based on track geometry measurement often show poor correlation with the actual vehicle response running on the track. Alternative methods based on vehicle response have been an emerging research focus, aiming to overcome the shortcomings of geometry-measurement-based methods. Track irregularity assessment based on vehicle acceleration works by establishing the correlation between vehicle response (dynamic output) to track irregularities (dynamic input). Studies done so far have focused on analysing the different types of track irregularities separately. Combined track irregularities have yet to become a focus of study, despite possible cross correlation affecting vehicle response. The present work aims to examine the use of machine learning methods to analyse combined track irregularities based on vehicle response. Rail vehicle response data generated through dynamic multibody simulation is filtered and processed into datasets of six features comprising standard deviation and peak value of carbody lateral acceleration, vertical acceleration, and roll acceleration per section of 100 metres. For machine learning training purpose, track irregularities dataset containing standard deviation and peak value of longitudinal level, alignment level, and cross level irregularities is also constructed. Machine learning classification is done with all six features as the predictors and both single-type irregularity class label and combined class label as the response, separately. Five classification algorithms are used: Decision tree, linear SVM, logistic regression, kNN, and random forest. Machine learning classification model training result shows high validation accuracy value across all classification cases and algorithms. Combined classification test result also presents high accuracy value. Further simulations are done with variance in rail vehicle speed, carbody mass, and wheel-rail friction coefficient. Previously trained combined classification model can predict the track irregularities condition in those condition with reasonable accuracy. This shows that the classification model is not overly sensitive to change in operational parameter, suggesting a high probability of success for use with ontrack measurement data. Track irregularities classification result is interpreted using Shapley additive explanation (SHAP). The Shapley explanation of the analysed track section indicates the most important predictors that explain the classification result. This indication can be used to analyse the type of track irregularities that is in unacceptable condition on that track section.