Performance comparison of machine learning algorithms for condition monitoring of tapered roller bearings

This paper investigated the implementation of machine learning algorithms for health monitoring and fault detection of tapered roller bearings (TRBs) (30205 J2/Q, 30206 J2/Q and 30207 J2/Q). Three defect models were considered: inner race defect, outer race defect and roller defect, in addition to data from the healthy bearings condition. An L27 orthogonal array design was used to generate a comprehensive dataset for each defect model, considering various operational parameters such as load, unbalance, defect type, bearing type and speed. Kurtosis was extracted as the sole feature from the vibration signals for fault classification. Several machine learning models, including artificial neural network (ANN), decision tree, support vector machine (SVM), random forest, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), gradient boosting and categorical boosting (CatBoost), were employed to predict fault severity. The results show that the ANN model accurately predicts faults based on the kurtosis metric. This study illustrates the capability of machine learning, particularly ANN, in enhancing the predictive maintenance strategies for TRBs, thereby enabling early fault detection under varying operational conditions.