A Machine Learning with an Inlier/Outlier Separation Approach for the Prediction of Wagon Maintenance Times

Time spent in wagons maintenance consumes a significant part of a rail freight company's budget. Thus, knowing how much time it is going to be spent in a maintenance procedure is critical for their management and planning. A common approach used to predict these time expenditures is the so called chronoanalysis. Despite their wide spread use, they may be inaccurate in some scenarios. Thus, in this paper, we try to replace it with machine leaning models which did not work at first. Then we propose a methodology that uses the chronoanalysis to divide the maintenance procedures into outliers and inliers. Hence, we were able to create independent models for each class. With this approach, the average mean absolute error was reduced from about 6 man-hour to a little above 2 man-hours. The best tested configuration presented an average mean absolute error of 0.417 man-hours compared with a 4.490 man-hours from the chronoanalysis.