A Vibration-Based Feature Selection Approach for Intelligent Diagnosis of Multiple Fault Conditions

Resumo

Three-phase induction motors play a vital role in industrial settings due to their robustness and versatility. However, they are subject to failure modes that compromise operational reliability, such as misalignment, unbalance, and bearing defects. Vibration signal analysis has emerged as an effective tool for diagnosing such conditions, enabling early anomaly detection and the implementation of predictive maintenance strategies. This study proposes an approach based on the selection of variables with high discriminative power to form reduced and representative feature vectors. The selection process combined correlation analysis with class-wise feature importance assessment using the Random Forest One-vs-Rest strategy. For each class, features contributing to the highest cumulative importance were retained, ensuring the preservation of the most relevant attributes for classification. This strategy effectively reduced data dimensionality without degrading model performance. To validate the effectiveness of the proposed selection, three widely used machine learning algorithms were employed: Random Forest, Support Vector Machine (SVM) with polynomial kernel, and XGBoost. These models were trained on the selected feature subsets and evaluated using performance metrics such as accuracy, precision, recall, F1-score, and ROC curves. Results demonstrated strong performance across all models, with SVM achieving the best overall results. Furthermore, model behavior varied across classes, suggesting that different algorithms may be better suited to specific fault patterns. The findings indicate that targeted feature selection combined with appropriate machine learning techniques forms an effective strategy for intelligent fault diagnosis in electric motors. The proposed methodology proves promising for industrial applications, offering a precise, scalable solution with lower computational cost and contributing to improved asset reliability and availability.