Detection of induction motor faults by envelopes of higher current harmonics

Abstract

Detection of broken rotor bars in induction motors based on sideband components near the fundamental stator current frequency often yields unreliable results under variable operating conditions due to masking effects and the dominant amplitude of the fundamental component. This study proposes a robust fault detection method which uses higher-order current harmonics, which demonstrate significantly reduced susceptibility to load variations and other masking phenomena. As principal contribution this study demonstrates that a convolutional neural network trained exclusively on data derived from a single operating condition achieves strong generalisation across a wide spectrum of unsteady operating regimes, spanning substantially different supply frequencies and load levels, with consistently high classification accuracy. Signal processing involves bandpass filtering around the fifth and seventh harmonic frequencies, envelope extraction via the Hilbert transform, continuous wavelet transform to generate time-frequency scalograms, and data augmentation through vertical shifts to simulate variations in defect-related frequency components under changing slip conditions. Scalograms corresponding to the fifth and seventh harmonics are treated as independent training samples, effectively expanding the dataset while leveraging physically redundant fault information. This single-point training paradigm directly addresses a critical industrial limitation, thereby enabling deployment of data-driven condition monitoring systems even under severe training data constraints.