Machine learning technique for damage detection of rails on steel railroad bridges subjected to moving train load

Abstract

Rail is one of the key elements of the railway system, and its role is to transmit the wheel load to the track bed and guide the train cars along the track. Rail is susceptible to rolling contact fatigue and wear due to being repeatedly subjected to the moving load of the train. This can eventually result in broken-rail damage and train derailment, which if happens on a railroad bridge, it can severely damage the bridge, such as the structural failure of the Tempe Town Lake steel railroad bridge in July 2020 that costed $11 million to repair. Therefore, early detection of defects in rail-bridge system may prevent a critical accident with irreversible damage. The objective of this paper is to use classification-based machine learning techniques to detect broken-rail damage in an open-deck railroad bridge by measuring its acceleration response under the moving load of the train for different speeds. For this purpose, the two-dimensional Finite Element (2D FE) model of a given railroad bridge is created using OpenSEESPy package, which is a Python-3 interpreter of OpenSEES. The changes in the acceleration response due to the damaged rail compared to the undamaged (healthy) rail are characterized by using the Hilbert-Huang Transform in both the time and frequency domains and quantified by defining energy and phase damage indices. The data collected from the 2D FE model are used to train and test several machine learning (ML) classifiers including the Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Decision Tree (DT) algorithms. The results from the data-analytic study show an acceptable level of precision of these classifiers in identifying the damage to the rail-bridge system.