PREDICTING AXIAL STRESS IN CONTINUOUS WELDED RAILS USING MACHINE LEARNING

Abstract

Buckling is a major contributor to derailments on continuous welded rails (CWRs) and is caused by extreme compression due to high temperatures. Thus, the determination of the axial stress is necessary to gauge when buckling may occur. As many techniques are often invasive, this study focuses on the latest advancements of an ongoing noninvasive technique based on the use of low-frequency vibrations using machine learning (ML). Based on the theory seen in guitar strings, the relationship between the frequency of vibration and tension/compression is used to associate spectral features to axial stress. This in turn enables the determination of the rail neutral temperature (RNT), which is the point at which the net longitudinal force is zero. As the RNT is not only a function of the stress but also varying boundary conditions associated with the ballast, fasteners, and ties, the RNT can change frequently over time. This stresses the need for a system that is also flexible to this variability. In order to capture some of this, field data was captured twice in Pueblo Colorado using several accelerometers where vibrations were induced with an instrumented hammer. Both a wood-tie and concrete-tie rail section were experimented on with the latter having a 5o curve. The accelerometers consisted of both wired and wireless sensors, with the wireless counterpart being used to prove the replacement of the wired in the second field test. Using the power spectral densities (PSDs) associated with the lateral and vertical direction as input to an artificial neural network (ANN), the RNTs were predicted and compared to those determined by an independent third party. Our predictions showed very good agreement with the RNT captured by conventional strain-gage rosettes. In this paper, an analysis of our two field tests as well as the impact of boundary conditions data on ML predictions is explored.