Small-sample Permanent Magnet Synchronous Motor Intern-turn Short Circuit Fault Diagnosis based on Strong Duality Theory

Inter turn short circuit (ITSC) faults, as common faults in permanent magnet synchronous motor (PMSM), have received increasing concern. This paper investigates the challenges faced by generative adversarial neural (GAN) networks in addressing the issue of small sample during the diagnostic process of PMSM ITSC, specifically focusing on the generator's gradient vanishing problem and the discriminator's inability to smoothly measure distances between different distributions. To address this challenge, this paper firstly proposed an improved duality gradient generative adversarial neural network (DGGAN) technology, which optimizes the loss function of the original GAN technology through strong duality theory to smoothly estimate distribution distances. Secondly, we utilize gradient penalty term to optimize the Lipschitz constraint process in strong duality theory, thereby avoiding the issue of low output convergence caused by enforced truncation constraints. Finally, to validate the effectiveness of DGGAN, we designed a real PMSM ITSC experiment to generate real world fault data. Then we used a combination of this real world data and GPGAN-generated pseudo-real data to train a deep diagnostic network model. Experimental validation shows that this method achieves a classification accuracy of up to 95.9% on real world test dataset.