Online Wheel Condition Detection Using Standard and Heteroscedastic Gaussian Process Learnings

Abstract

The vibration-based wheel condition monitoring may suffer from measurement noises as well as stochastic wheel/rail dynamics; in this regard, statistical methods are believed to be more promising in dealing with these variabilities in monitoring data. This paper proposes two state-of-the-art Bayesian machine learning approaches for online wheel condition assessment based on dynamic strain monitoring of rail tracks. In this first approach, wheel quality-sensitive ingredients are firstly extracted from the original signals and their Fourier amplitude spectra are normalized to characterize informative features for assessment of wheel qualities. A statistical model is then formulated by standard Gaussian process regression for representation of wheels under normal condition such that wheel conditions can be quantitatively and qualitatively evaluated with the aid of Bayes factor once new monitoring data are available. In the second approach, the uncorrelated and uniform assumption on measurement noises in standard Gaussian process is removed and heteroscedastic Gaussian process allows error variance to vary over time, resulting in more robust and reliable diagnosis of wheel qualities when heteroscedasticity is manifested in the data. The two proposed approaches are examined by using online strain monitoring data collected by an optic fiber sensory system.