A Bayesian Machine Learning Approach for Online Wheel Condition Detection Using Track-side Monitoring

Abstract

Online wheel condition monitoring can suffer from the stochastic wheel/rail dynamics and measurement noises. This paper aims to develop a Bayesian statistical approach for probabilistic assessment of wheel conditions using track-side monitoring. In this approach, the wheel quality-related components are first extracted from monitoring data and their Fourier amplitude spectra are normalized to obtain a set of cumulative distribution functions that characterize wheel quality information. Then a data-driven reference model is established by means of sparse Bayesian learning for modelling these characteristic functions for healthy wheels. Bayes factor is finally employed to discriminate the new observations from the reference model, with which a quantitative evaluation of wheel qualities is achieved in real time. To validate the feasibility and effectiveness, the proposed approach is examined by using strain monitoring data of rail bending acquired from a track-side monitoring system based on optical fiber sensors.