Machine learning-based feature-driven model generation and evaluation for multi-fault bearing diagnosis using XGBoost and time-domain statistical features of vibration data

Abstract

Accurate and timely detection of bearing faults is critical for ensuring the reliability and safety of rotating machinery in industrial environments. This study investigates the application of machine learning (ML) techniques for classifying ten different types of bearing fault using features extracted from vibration signals. Among the various classifiers evaluated, the extreme gradient boosting (XGBoost) algorithm demonstrates superior performance. Through a combination of standard signal preprocessing, hyperparameter optimisation via GridSearchCV and comprehensive model evaluation using accuracy, F1 score and confusion matrix analysis, the optimised XGBoost model achieves a high classification accuracy of 95.48%. Notably, fault types such as IR_007_1, IR_014_1 and OR_007_6_1 are classified with perfect precision and recall. While minor misclassifications occur, primarily between ball and outer race (OR) faults due to overlapping signal characteristics, the overall confusion matrix underscores the strong generalisation and class discrimination capabilities of the model. Additionally, feature importance analysis highlights kurtosis, standard deviation (SD) and mean as the most influential statistical descriptors, with kurtosis emerging as the dominant factor for distinguishing impulsive fault characteristics. These findings affirm the efficacy of XGBoost in bearing fault classification and emphasise the diagnostic value of carefully selected time-domain features. The results suggest strong potential for deploying such models in real-time condition monitoring and predictive maintenance systems. Future research may explore integrating deep learning or hybrid ensemble methods to further enhance diagnostic accuracy in more complex or noisy fault scenarios.