Machine Learning for the Nondestructive Prediction of Neutral Temperature in Continuous Welded Rails

Abstract

ABSTRACT This paper describes the application of machine learning (ML) in the framework of a data-driven nondestructive evaluation (NDE) method to estimate the rail neutral temperature (RNT) of continuous welded rails (CWR). The method consists of triggering vibration of the rail of interest and extracting the power spectral densities (PSDs) of the accelerations associated with the lowest modes of vibration. The PSDs then become the input of an ML algorithm trained to associate the PSD to longitudinal stress and then RNT. In the study presented in this article, the proposed NDE method was tested on a tangent track on wood cross-ties. Vibrations were induced with a hammer and detected with several wireless and wired accelerometers. The PSDs across the 0–700 Hz range were extracted from the time-series. These densities in both the lateral and vertical directions constituted part of the input of an artificial neural network trained and tested with experimental data. The predicted neutral temperatures showed very good agreement with the RNT estimated by an independent party and based on conventional strain-gage rosettes.