Predição de falha em motor diesel de locomotiva baseada na análise do óleo lubrificante por meio de técnicas de aprendizado de máquina

Abstract

Diesel-electric locomotives are the most used in Brazilian freight railroads, with the diesel engine being an indispensable part of its operation. If the diesel engine fails, the train pulled by the locomotives will not circulate, which may lead to non-compliance with transport plans and costs for the railways. This dissertation aimed to investigate two machine learning algorithms: the Artificial Neural Network (RNA) and the XGBoost Decision Tree, in the prediction of locomotive diesel engine failures through real data acquired on the Vitória-Minas Railway, located in the Southeast Region of Brazil. The input data included the results of lubricating oil analyzes, such as those of spectrometry and physical-chemical tests. The history of engine failures that have already occurred, the fuel consumption of the locomotives and the movement of engines between locomotives were also used as input. Correlation, normalization, and data balancing analyzes were performed. In addition, for both prediction methods, k-fold cross validation was applied. The classifiers were evaluated for accuracy, recall and classification indications in the confusion matrix. Both algorithms achieved good performance to predict failures in diesel engines, being able to assist the engine maintenance planner in locomotive workshops.