Predictive Maintenance System Using Machine Learning

MICHAEL R. MATTHEWS, TAPAN SONI, JOHN A. STRANAHAN, JOSHUA C. JACKSON, NICHOLAS LA SALA, and CRAIG WERT

For many types of equipment, maintenance needs to be performed opportunistically. For example, naval vessels may have small maintenance windows when they are docked. Therefore, it is in the best interest of the maintenance crew to be aware of whether a particular part is vulnerable for failure in the near future. In the past, preventative maintenance was performed based on manufacturers' specifications and guidelines. But in reality, a one-size-fits-all maintenance model does not account for individual variations in equipment. Our research is centered around using Machine Learning techniques to predict when an engine system is in need of repair. This can be done by creating a classifier using historical datasets. Such data can be collected in real time using onboard sensors which report on specific parameters about the equipment in question. To predict the maintenance window, we researched several supervised classifiers such as Support Vector Machine, Decision Tree, Gradient Descent, and K-Neighbor. Ultimately we selected the K-Neighbor classifying algorithm which bases its decision on the nearest known data points. Then,
after normalizing the current engine dataset and splitting it into a training and validating set, we can make accurate predictions over all the rows of the data and return either a 'Yes, the engine needs repair', or 'No, all is good' depending on the average of all the predictions.