Latest Advancements on a Vibration-Based Monitoring Approach for Continuous Welded Rails

This paper describes a monitoring/inspection technique for the estimation of longitudinal stress in continuous welded rails (CWR) to infer the rail neutral temperature (RNT), i.e. the temperature at which the net longitudinal force in the rail is zero. The technique relies on the extraction of the spectral densities of the lower modes of vibration of the rail of interest and the application of machine learning algorithm (MLA) to predict the RNT. In the study presented in this article a custom neural network was developed using hyperparameter search optimization techniques. The frequencies of vibration extracted from the power spectral densities (PSD) of the vibration data were used as input parameters to the MLA. To further determine the important frequencies within the entire PSD, the minimum redundancy-maximum relevance (mRMR) algorithm was utilized. This resulted in several sets of reduced input frequencies where they were tested on various ML algorithms. The results of two field experiments demonstrated that the neural network can capture the variability in boundary conditions present alongside predicting the RNT. Furthermore, the models using reduced feature sets under mRMR were able to sufficiently predict the RNT with minimal impact on performance, confirming that the PSDs contain redundant information that can be removed in future implementations upon further data collection.