A Novel Method for Bearing Fault Diagnosis Based on Novel Feature Sets with Machine Learning Technique

Abstract Bearings often experience small and medium raceway damage due to operating and loading conditions, which induces abnormal dynamic behavior. In this study, a dynamic model of the bearing system with various conditions and bearing faults is developed based on experimental investigations, Extreme Machine Learning (EML), and supervised machine learning K-Nearest Neighbors (KNN). The effects of defects on system dynamic response and the damage vibration of the bearing are investigated through simulation. Experiments verify the typical dynamic characteristics. The fundamental bearing characteristics frequencies and statistical features withdrawn from a vibration response are utilized for fault identification using a machine learning algorithm. Bearing characteristics, frequencies, and statistical features were applied to both proposed approaches and compared regarding their prediction quality. The result shows that the EML performs better than the KNN in terms of precision of fault recognition by up to 99%. This work provides valuable insights for operation, maintenance, and early fault warning related to bearings.