Permanent Magnet Synchronous Motor Fault Detection System Based on Transfer Learning Method

Permanent magnet synchronous motors (PMSMs) are increasingly used in industrial and commercial applications. Due to their popularity, the fault detection of these machines constitutes an extremely important issue. Currently used diagnostic tools are mostly developed based on deep or shallow neural structures, and training processes require large learning data packets. This fact results in the need for interference in motor construction to obtain measurable fault symptoms for NN training. To limit physical modelling of damages, mathematical models based on finite element methods (FEM) are mostly used. Nevertheless, the development of neuronal detectors based on simulation results does not provide high accuracy and a short time to implementation in case of changing the diagnostic task. To solve that limitation, the PMSM fault detection system based on the convolutional neural network (CNN) trained according to the transfer learning (TL) method was proposed in the article. The research aims to show the possibility of detecting PMSM faults (partial demagnetization and stator winding fault) in steady and transient states. The CNN was trained using only the phase current signals coming from the FEM model. The experimental verification was carried out on the PMSM motor during changes in drive operating conditions. The results of the experimental research carried out on a specially designed PMSM show the impressive capability of the developed CNN-based diagnostic system obtained using the transfer learning method.