An Artificial Neural Network Model for Maintenance Planning of Metro Trains

Abstract

In urban transportation, trains have an increasingly important place due to the increase in the number of passengers. Meeting the number of passengers is directly related to the number of trains operated on a line. Thus, the frequency of operation of trains affects the level of wear of the equipment. This makes train maintenance more important. Equipment faults are the basis for train maintenance. However, the fault times of the equipment which are unknown causes uncertainty in the maintenance activities and plans. This uncertainty results from many factors that affect the faults of the train. If historical maintenance data, fault data, and factors affecting the faults are known, effective use of resources (time, cost and personnel, etc.) is provided and uncertainty is eliminated. In this study, firstly, maintenance data in Ankara Metro between 2017 and 2018 is examined and the factors affecting equipment faults are evaluated with expert opinion. Artificial Neural Network (ANN) model is created with the data set and this data set along with the factors affecting each the equipment fault according to the type of equipment. In the ANN model, 5 factors (Equipment Type, Preventive Maintenance Frequency, Material Quality, Life Cycle, Line Status) affecting the faults of the equipment is determined as inputs and the number of failures as outputs. The mean absolute percent error (MAPE) value is found as 11%, and the mean square error value (MSE) is 0.0028229 in the training and test stages of ANN. Then, the frequency of fault is found according to the equipment fault and a 10-week maintenance planning is applied. The results are compared with current maintenance planning. As a result of the applied maintenance planning, the average number of faults of the trains decreases by 27%, uninterrupted service rate increases by 40% and heavy maintenance errors are also prevented. Fault removal times resulted in a 10% improvement. The results showed that ANN models could be used effectively in fault prediction and maintenance planning with rail system multiple types of equipment. In the literature, there is no study that implements maintenance planning with an ANN model where all train equipments and factors affecting the failure are evaluated together. This study is the first in the field of rail systems maintenance in the literature and will be a reference for future studies.