Few-Shot Learning With Residual Current for EV Inverter Fault Diagnosis of EV Powertrain

The data-driven approach plays a critical role in the reliability of the permanent magnet synchronous motor (PMSM) drive for electric vehicles (EVs). Generally, data limitations and the ability of algorithms to extract fault features significantly affect the performance of the fault diagnosis. It is challenging to obtain the desired number and quality of samples due to the huge cost of obtaining fault data for motor drive systems under complex operating conditions. Considering the drawbacks of data-driven methods for fault diagnosis including hardware costs and data uncertainty, a combined model and data-driven few-shot learning network was proposed to detect voltage source inverter (VSI) open-circuit faults in PMSM in this article. The proposed method utilized the residuals obtained from the observer data and the actual measurement data being used as raw samples to get rid of the effects of harmonics and thus improve the quality of the input data to reduce diagnostic uncertainty. Then, an attention-based vision transformer (ViT) model is applied to explore the association of global features of the input samples rather than local features for the input sample by signal dimension conversion, in combination with the Siamese network framework. The experimental results verify that the proposed method reached an impressive diagnostic accuracy rate of 92.85% which has a better performance compared with classical methods under varying faulty load conditions.