Signal processing and graph-based semi-supervised learning-based fault diagnosis for direct online induction motors

Abstract

In this thesis, fault diagnosis approaches for direct online induction motors are proposed using signal processing and graph-based semi-supervised learning (GSSL). These approaches are developed using experimental data obtained in the lab for two identical 0.25 HP three-phase squirrel-cage induction motors. Various electrical and mechanical single- and multi-faults are applied to each motor during experiments. Three-phase stator currents and three-dimensional vibration signals are recorded simultaneously in each experiment. In this thesis, Power Spectral Density (PSD)-based stator current amplitude spectrum analysis and one-dimensional Complex Continuous Wavelet Transform (CWT)-based stator current time-scale spectrum analysis are employed to detect broken rotor bar (BRB) faults. An effective single- and multi-fault diagnosis approach is developed using GSSL, where discrete wavelet transform (DWT) is applied to extract features from experimental stator current and vibration data. Three GSSL algorithms (Local and global consistency (LGC), Gaussian field and harmonic functions (GFHF), and greedy-gradient max-cut (GGMC)) are adopted and compared in this study. To enable machine learning for untested motor operating conditions, mathematical equations to calculate features for untested conditions are developed using curve fitting and features obtained from experimental data of tested conditions.