Automating predictive maintenance using oil analysis and machine learning
Sarah Keartland, Terence L. van Zyl
Abstract
Predictive maintenance aims to reduce costly and time consuming repairs, and also avoid unnecessary activities by proposing a maintenance strategy that is informed by machine condition monitoring. The majority of mechanical systems are oil lubricated, therefore oil analysis provides a rich source of machine condition data for many mechanical systems. This research investigates the use of random forests, feed-forward neural networks and logistic regression models trained using oil analysis data for classifying machine conditions. The RF model outperformed the other classifiers for all machine conditions. The interpretation of the feature importance for the RF models were found to be consistent with industry knowledge, demonstrating the potential use of RF as a diagnostic tool in predictive maintenance.