Classification of Multiple Incipient Faults in Permanent Magnet Synchronous Motors under Variable Speed Conditions Using Transfer Learning

This paper presents a classification of multiple faults in permanent magnet synchronous motors at low severity levels under variable speed conditions using a transfer learning-based approach with pre-trained models. The classifier demonstrates the effectiveness of stray flux and current sensors in detecting multiple low-severity faults, specifically local demagnetization and 5% inter-turn short circuit faults. The measured flux data from two locations are converted into spectrogram images using a short-time Fourier transform. A preprocessing method for sequential spectro-images has been developed to mitigate high variance and bias error in training. Within this framework, five pre-trained models from ImageNet are utilized to validate the accuracy and performance at two different sensor locations: ResNet50, VGG16, Inception, Xception, and MobileNet The suggested method is validated with data from an in-house experimental setup.