A Comparative Study of Wavelet and EMD Methods for Radar-Based Detection of Rotational Imbalance in Induction Motors

This paper presents a radar-based methodology for detecting rotational imbalance in machinery. The approach employs multi-resolution signal processing techniques, specifically comparing Empirical Mode Decomposition (EMD) and Wavelet Decomposition (WD), for feature extraction from radar-acquired signals. Both methods generated five levels of sub-components. Time-domain features were extracted individually from each sub-component, and the resultant feature vectors from all levels were concatenated to form the final input for machine learning classifiers. WD was evaluated using three distinct mother wavelet functions (Haar, Coiflet-2, and Symlet-4), with the performance of each benchmarked against EMD. The comparative analysis rigorously assessed classification metrics (accuracy, precision, recall, F1-score) alongside computational efficiency metrics (model size, training and prediction latency). Results demonstrate that wavelet-derived features consistently outperformed EMD across all classifiers. The Haar-Random Forest combination achieved the highest classification accuracy (95.28%). Support Vector Machines paired with Haar features offered an optimal accuracy-latency trade-off (approximately 219 ms training time), while Decision Trees delivered ultra-fast inference (0.6 ms). These findings validate the practical viability and effectiveness of the proposed radar-based vibration detection framework for industrial condition monitoring systems.