This study presents a comparative evaluation of feature selection and dimensionality-reduction techniques for vibration-based fault diagnosis of induction motors using machine learning. Three approaches are systematically analyzed: Minimum Redundancy Maximum Relevance, Principal Component Analysis, and their combined application. A broad benchmark of thirty-three classification models was conducted using the MATLAB Classifier Learner App to assess classification accuracy, model complexity, and prediction speed. The use MATLAB environment ensured reproducibility and facilitated systematic comparison across different algorithms. The results show that PCA in combination with a Quadratic Support Vector Machine achieves the highest diagnostic accuracy (99.8 %). Meanwhile, mRMR paired with a Narrow Neural Network offers an optimal balance between accuracy (99.5 %) and computational efficiency, delivering a prediction speed nearly 7.5 times faster than the leading PCA model. The combined mRMR?PCA approach demonstrates reduced effectiveness, indicating limited benefit from sequential feature selection and extraction for this dataset. The proposed methodology highlights the practical value of integrating vibration data with machine learning techniques.
Stojiljkovic et al. (Thu,) studied this question.