Nondestructive Estimation of Neutral Temperature in Rails: A Comparative Study of Machine Learning Strategies

Abstract

This paper presents the latest findings of a nondestructive evaluation technique currently under development at the University of Pittsburgh to determine the rail neutral temperature (RNT) in continuous welded rails. The technique is based on the extraction of relevant features from rail vibrations and the use of machine learning (ML) to associate these features to the longitudinal stress of the rail of interest. The features contain the spectral information of the vibrations and are pooled together by frequency domain decomposition for input to ML algorithms. Minimum redundancy–maximum relevance and neighboring component analysis are used to identify relevant features to reduce the size of the input vector. In addition, seven algorithms were considered to identify the most accurate model for neutral temperature with respect to the ground truth RNT measured with a strain-gage rosette. The data used in this study were collected from a 5° curved rail on concrete ties. The vibrations were triggered with a hammer and recorded with a few wireless and wired accelerometers attached on the railhead. The results showed that the Gaussian process regressor performs best, and as few as 20 frequencies can be used to predict the RNT with sufficient accuracy.