Rail neutral temperature prediction via Local Resonance and Gaussian process regression

The rail neutral temperature (RNT), where there is zero stress in continuous welded rail (CWR), is an important measure for rail maintenance. Due to environmental and operational factors, including ambient temperature, supporting track structure, curvature, and maintenance activities, it is very challenging to measure and monitor RNT without the reference at a known stress state in a non-destructive and non-disruptive manner. To address the intractable task, this work explores using rail local resonances and Gaussian process regression to predict RNT. The local resonance frequencies as well as rail temperatures were served as the features for the inputs. The Gaussian kernel was used for computing the variance matrix while Laplacian scores were used as the method for feature selections. It was decided to use three most important resonance frequencies plus the rail temperature for the feature vector to balance accuracy and computation efficiency. The prediction results suggest the Gaussian process regression could not only give accurate RNT predictions but also indicate the prediction variance, where common machine learning methods do not have.