A Machine-Learning Approach for the Assessment of Quantitative Changes in the Tractor Diesel-Engine Oil During Exploitation

Abstract

To evaluate the potential of a machine-learning approach in the assessment of quantitative changes in the tractor diesel-engine oil during exploitation, this study aspired to propose a machine-learning regression method to reduce the frequency of expensive and time–consuming engine oil sampling. The input engine-oil datasets with fresh engine oil (Samples A) and with the engine oil subsequent to 250 working hours (Samples B) were sampled for twelve elements in a two–year exploitation study at the Belje company. The field data collection was performed having deployed six heavy, four–wheel-drive FENDT 930 Vario agricultural tractors, each monitored for 1,500 working hours, during which an engine oil was sampled every 250 working hours. The evaluated machine-learning prediction methods, based on a tenfold cross–validation, achieved a moderately high prediction accuracy, with a slightly higher coefficient of determination (R2), in the range of 0.51–0.73, for the Samples B, than those in the range of 0.49–0.64 for the Samples A. These results strongly suggest that none of the machine learning methods constantly achieved high prediction accuracy and that the selection of optimal machine-learning models should be mandatory, having also confirmed a high potential of machine-learning methods in the detection of quantitative changes in tractor diesel-engine oil during exploitation.