An Interpretable Machine Learning Approach For Fault Classification in Bearing Systems

This paper presents an interpretable machine learning approach to classify different types of faults in ball bearing systems with high accuracy. These methods are based on ensemble methods (Random Forest, Gradient Boosting, and AdaBoost) which use data from the fan system collected by vibration monitors mounted on three axes. The proposed technique needs statistical feature inputs to classify different types of faults. We train models and explain their output using SHAP (SHapley Additive exPlanations). In the rigorous numerical study, we point out a suitable explanation for black-box predictive models trained on vibration data. Our model achieves 98 percent classification accuracy when classifying all fault types.