Hybrid Approach for Bearing Fault Diagnosis in Induction Motors Based on Isolation Forest and Ensemble Learning

Rolling element bearings are fundamental parts of rotating machinery, and their sudden breakdown may cause abnormal vibrations, an unplanned production halt, and higher maintenance costs. To address this problem, this paper presents a hybrid data-driven approach for the automatic diagnosis and classification of bearing faults using vibration signals from the IMS dataset. The proposed method first divides the raw vibration signals into fixed-length windows and computes seven statistical features to characterise the signals. An unsupervised Isolation Forest technique is then used to detect anomalous signal segments, providing an early warning of potential faults. Thus, it does not require prior knowledge of the normal condition for fault detection. Subsequently, the detected anomalies are classified using two ensemble-based supervised learning models: Extra Trees and XGBoost. The experimental results indicate that the tree-based ensemble classifiers outperform the other models tested. Specifically, XGBoost achieves an F1 score of 97. 41%, whereas Extra Trees achieves an F1 score of nearly 97%, indicating its high potential for accurately detecting different types of bearing faults. The results confirm that the proposed hybrid two-stage model is an efficient and reliable tool for bearing fault diagnosis, supporting early fault detection, and enhancing system reliability in industrial settings.