Integrated Machine Learning Approaches for Comprehensive Bearing Health Monitoring and Fault Classification Using Multi-Sensory Data

Abstract

Modern industries heavily rely on machines equipped with rolling-element (RE) bearings. However, these machines face substantial risks due to potential bearing faults, where even minor defects can lead to catastrophic failures. Shockingly, statistics reveal that up to 40-51%of induction motor failures can be attributed to bearing damage. Early fault detection through Condition Monitoring is therefore crucial. Over the last two decades, various machine learning (ML) techniques have been explored to detect defects in rolling element bearings. This research project focuses on optimizing ball-bearing fault detection through diverse ML techniques, supported by comprehensive experimental work. The experimental work entails multiple steps, including the preparation of a varied set of bearings and the enhancement of equipment with advanced sensors. These experiments resulted in a rich dataset comprising vibrations, currents, and sound – vital for ML model training. The experiments spanned two different machines, incorporating variations in speed and applied forces. Analysis of the experimental data unveiled the significant influence of defect shapes on bearing responses. Notably, at lower speeds, defect shapes prominently affected vibrations, with rectangular defects displaying logarithmic growth while circular defects exhibited exponential behavior. With increasing speed, the behavior of bearings with different defect shapes tended to converge. Current emerged as a robust choice for fault detection, maintaining consistent behavior regardless of defect size and shape, while sound closely resembled vibrations, with slight variations in the case of circular defective bearings. The second part of this research is dedicated to developing ML fault detection models. Three models, utilizing ensemble techniques, were created. These models, employing Decision Trees (DT), Random Forest (RF), and XGBoost, achieved prediction accuracies of 82%, 91%, and 92%, respectively. A feature importance analysis identified CRSF and SF as dominant parameters. Furthermore, a sound-to-vibration transformation ML model was introduced. This model, built on a 1D Operational U-Net (Op-UNet) framework, is capable of synthesizing realistic vibration signals from sound measurements across different working conditions, fault types, and severities achieving a striking minimum accuracy of 97%. The models and datasets presented in this research signify a significant advancement in bearing health condition monitoring.