| dc.description.abstract | Keeping large-scale transportation infrastructure networks, such as railway networks, operational under all conditions is one of the major challenges today. The budgetary constraints for maintenance purposes and the network dimension are two of the main factors that make the management of a transportation network such a challenging task. Accordingly, aiming to assist the management of a transportation network, a data-driven model is proposed for stability condition prediction of embankment slopes. For such a purpose, the highly flexible learning capabilities of artificial neural networks (ANN) and support vector machines (SVM) were used to fit data-driven models for earthwork hazard category (EHC) prediction. Moreover, the data-driven models were created using visual information that is easy to collect during routine inspections. The proposed models were addressed following two different data modeling strategies: nominal classification and regression. Moreover, to overcome the problem of imbalanced data (since typically good conditions are much common than bad ones), three training sampling approaches were explored: no resampling, synthetic minority oversampling technique (SMOTE), and oversampling. The achieved modeling results are presented and discussed, comparing the predictive performance of ANN and SVM algorithms, as well as the effect of the sampling approaches. A comparison between nominal classification and regression strategies was also carried out. Moreover, aiming at a better understanding of the proposed data-driven models, a detailed sensitivity analysis was applied, allowing the quantification of the relative importance of each model input, as well as measuring their global effect on the prediction of embankment stability conditions. | |
