Estimating Rail Neutral Temperature using Local Resonance and Probabilistic Machine Learning

Abstract

The continuous welded rail (CWR) has been widely used in modern railways due to their support to high transport speed while requiring less maintenance compared with jointed track. However, CWR is susceptible to internal stress because of restrained free thermal expansion and contraction in its axial direction. During summer, the excessive heat on the rail will cause expansion of the rail, but it is constrained, leading to a huge axial force. If the axial force is excessively large, thermal buckling could be triggered depending on track conditions. In fact, the sense and magnitude of the built-up rail axial stress depend on the rail temperature relative to the set rail neutral temperature (RNT). Estimating RNT without taking baseline measurements has been a long-standing challenge for the railway engineering community. In this study, we present the most recent advancement in RNT estimation using intrinsic local resonances in rails. This study demonstrates that frequencies of these local resonances, including zero-group velocity and cutoff frequency resonances whose energy are trapped locally close to the load zone, are insensitive to the presence of rail supports and sensitive to rail temperature and axial loads, which has great potential for RNT estimation. The data collection system can consistently extract rail local resonances from a fully instrumented revenue-service site, where strain gauges and thermocouples were attached and calibrated during the track construction and can supply the ground truth of RNT and rail thermal forces. Probabilistic machine learning algorithms are developed to predict RNT of the revenue-service site using rail temperature and frequencies of rail local resonances, both of which are directly measurable. The performance of the proposed machine learning framework is evaluated by comparing the predictions with the ground truth.