Wear fault diagnosis in journal bearings using vibration analysis and AI

In this study, machine learning models for identifying wear defects in journal bearings under various operating circumstances are compared. Healthy, low, and high wear states were simulated using an experimental test rig. Information Gain, Gain Ratio, Gini Index, and other feature selection techniques were used to analyze vibration signals from precision sensors. Relevance was used to rank features like Root Mean Square, Peak-to-Peak, and Peak. Using accuracy, precision, recall, F1-score, specificity, and log loss, models from Gradient Boosting, AdaBoost, Random Forest, k-nearest Neighbours, Support Vector Machine, and Decision Tree were assessed. Gradient Boosting performed best overall and had the highest accuracy (99.5%). Random Forest and AdaBoost also showed high classification accuracy. More features were beneficial for simpler models like k-Nearest Neighbours and Decision Tree. Random Forest and Gradient Boosting successfully decreased misclassification between related fault classes. The findings highlight the importance of selecting the appropriate models and features. Advanced feature engineering and parameter optimization could lead to further enhancements.