Data analytics and machine learning for railway track degradation: Using Bothnia Line track measurements for maintenance forecasting

Abstract: In this paper, a statistical method is developed to improve predictive maintenance on railway track. The problem tackled is being able to predict when the next maintenance event should take place to guarantee a certain track quality class. To solve the problem, The prediction is made using track measurement data exclusively, with no maintenance history to support the data analysis. The dataset consists of track measurements taken over eleven years and 170 kilometres on the Bothnia line in Northern Sweden. Different track degradation models and machine learning approaches are discussed and implemented. In the end, the tool developed was able to predict track degradation with an error within reasonable bounds of the typical maintenance limit. This will allow an operator to predict the recommended date for the next maintenance event at all locations using only historical track measurements as an input and little to no user intervention on the programme.