Enhanced diagnosis and monitoring of broken rotor bar faults in induction motors using a combined CEEMDAN-MLP approach

This paper presents a novel approach for diagnosing and monitoring Broken Rotor Bar (BRB) faults in induction motors through vibration signal analysis. The method integrates advanced signal processing techniques such as the Hilbert Huang Transform (HHT) with machine learning methods, specifically Multilayer Perceptron (MLP). The study initiates with an HHT application to identify fault-related harmonics, achieved through complete Empirical Ensemble Mode Decomposition with Adaptive Noise (CEEMDAN) of the vibration signal (Vx), producing intrinsic mode functions (IMFs). A statistical analysis, employing correlation coefficients (CC), facilitates the selection of relevant IMFs indicative of BRB faults. IMFs with CC values equal to or greater than 0.2, notably IMF1, IMF2, IMF3, and IMF4, appear informative. Following IMF selection, signal reconstruction ensues by incorporating these useful IMFs. After rebuilding the signal, we use global thresholding based on a statistical analysis that includes Root Mean Square (RMS) and Energy Coefficient (EC) calculations. The Signal Reconstruction Denoising (SRD) meets the criteria for selection. Spectral envelope analysis of SRD is then employed for BRB fault detection. The subsequent phase employs a Multi-Layer Perceptron (MLP) for BRB localization. Features utilized for training the MLP model include EC and various frequency components (fvb-, fvb+, 2fvb-, 2fvb+, 4fvb-, 4fvb+, 6fvb-, 6fvb+, 8fvb-, and 8fvb+). Results from MLP demonstrate exceptional performance, achieving a classification rate of 99.99%. The proposed CEEMDAN-MLP method exhibits robust efficiency, validated by experimental results, and offers promising prospects for BRB fault diagnosis and monitoring in induction motors.