A Few-Shot Open-Circuit Fault Diagnosis of F-Type Inverters Using CGAN-Based Vision Transformer

Abstract

Multilevel inverters (MLIs) are widely adopted in various industries due to their distinctive features. However, they are susceptible to higher failure rates due to the increased number of components. Deep learning (DL) models are widely used for accurately diagnosing faults in inverters because they effectively extract features automatically. These models work on the hypothesis of the availability of a sufficient number of samples to train the diagnostic models. However, obtaining enough sample data in engineering practice is difficult, especially in fault cases. Therefore, this article proposes a fault diagnosis scheme combining a conditional generative adversarial network (CGAN) and a vision transformer (ViT) for diagnosing open-circuit (OC) faults with few fault samples (few shots). First, the measured signals are converted to time–frequency images. Afterward, CGAN generates new 2-D sample images with data distributions similar to real samples. Finally, the improved ViT uses the original and generated samples to learn and extract local and global features with a multihead self-attention mechanism and classify the faults. The proposed scheme is validated using an experimental setup of the F-type inverter, and the results show that the suggested scheme outperforms the other conventional DL methods with an accuracy of 98.46% using only six samples per class.