PREDICTIVE MAINTENANCE OF BALL BEARINGS USING CONVOLUTIONAL NEURAL NETWORKS (CNN)

Abstract

To this day, numerous maintenance actions follow preventive and run-to-failure methods. In this work, it has been attempted to show the power of predictive maintenance (PdM) from vibration data using a machine learning technique implementing convolutional neural networks (CNN). Actual data was collected from six different bearings on a machine element fault analysis test rig. The bearings were of the ball bearing type, where one of them was healthy and the rest were faulty at the inner race, the outer race, the rolling element, or a combination of these three, and another had severe damage at either the rolling element or the rings. From the vibration signatures specific to the health status of the bearings, a powerful deep learning convolutional neural network model was built. The model was able to successfully classify the states of bearings with accuracy values ranging between 76.7 and 99.9% based on unseen data. The results indicate that this CNN model can be used as a diagnostic tool for undertaking maintenance operations.
KEYWORDS: CNN, predictive maintenance