Broken Rotor Bar Fault Diagnosis Using Spectrum Distortion for Data Augmentation

The problem of diagnostics of broken rotor bars in an induction (IM) motor based on its current signals is considered. The proposed fault detection technique is based on filtering the current signal in the region around the frequency 7f, where f is the power supply frequency, and further obtaining its envelope using the Hilbert Transform. Scalograms are constructed from the envelopes using a Continuous Wavelet Transform and form a training dataset for the Convolutional Neural Network (CNN) classifier. In real industrial conditions, we have a lack of defective data which is simulated in this paper. It is shown that a CNN model trained on normal data obtained under different motor operating conditions and defective data obtained under one operating mode will poorly detect a defect when changing the operating mode. The reason for this lies in the characteristic defect frequencies change with the change in the operating mode. A data augmentation method based on spectrum distortion is proposed to solve this problem. We distort the spectrum of a real defective signal taking into account its structure and create synthetic data that simulate missing defective signals with different operating conditions. The paper presents the experimental rig and dataset, and shows that the spectral distortion-based augmentation technique significantly improves the performance of the CNN model under conditions of insufficient defective training data.