A review on rail remaining useful life prediction

Abstract Rail is one of the most basic components of railway tracks. Accurate prediction of the remaining useful life of rail is important for ensuring operational safety, optimizing maintenance resource allocation and reducing operational costs. Nevertheless, research on the prediction of the remaining useful life of rail is limited, primarily due to the challenge of obtaining complete lifecycle data and the complexity of degradation mechanisms. This paper presents a review of the literature on prediction of the remaining useful life of rail, with a focused summary of the developmental trends, advantages and limitations of three predominant methodologies: mechanism analysis methods, data-driven methods, and mechanism analysis and data-driven fusion methods. The results show that the material wear loss model and crack propagation model are mostly used in the mechanism analysis methods, and the remaining useful life of the rail is predicted by modelling the degradation mechanism model. Despite its high predictive accuracy, the mechanism analysis model for rail crack and wear evolution has limitations, including complex modelling procedures and a narrow range of applicability. The data-driven method analyses the data monitored throughout the full lifecycle of the rails, mines the rail degradation law hidden behind the data, and predicts the remaining useful life of the rail through the statistical probability method and machine learning. Although this method can be rapidly deployed with a sufficient volume of high-quality monitoring data, its predictive accuracy is susceptible to data noise. The mechanism analysis and data-driven fusion methods simultaneously analyse the rail defects mechanism and monitoring data, and use the rail monitoring data to estimate the parameters of the mechanism degradation model, improving the accuracy of the prediction of the remaining useful life of the rail. This method requires a grasp of the rail deterioration mechanism and abundant monitoring data, thereby inheriting the strengths and limitations of the aforementioned two approaches. Finally, the challenges and potential research directions in the future are discussed.