Maintenance Applications of a Machine Learning Model of Rail Defects

Rolling contact fatigue is the first cause of rail degradation.The phenomenon is deemed random yet depending on many exogenous parameters.Numerical physical modeling has shown its limits and data modeling had to be introduced.In this paper, one proposes to show diverse applications of a hybrid model using data and physics in an industrial context.The data architecture linking physical model, random forest classification model and survival forest model is quickly exposed.Then some indicators are proposed to eval the model performance.Three applications are then developed based on this mixed-model, phenomenon understanding -closely linked with model validation by expert knowledge -, predictive maintenance scheduling by rail grinding, and prescriptive maintenance illustrated by fixing the time delay before the first visit by ultrasound engines.