Using DCGAN and WGAN-GP to Generate Artificial Thermal RGB Images for Induction Motors

The paper investigates the feasibility of using GANs to create realistic induction motor thermal RGB image datasets for multimodal condition-monitoring systems. Generating high-quality thermal images presents computational challenges, and in this study, two GAN frameworks, DCGAN and WGAN-GP, were used under different health conditions. Firstly, DCGAN was used on three conditions using various hyperparameters, but the results required further improvement. Secondly, WGAN-GP was used with an extensive training duration of 11 hours, utilising 10,000 epochs and a batch size of 64, targeting the inner fault dataset, which resulted in generating artificial images that closely resembled real images. This study highlights the impact of hyperparameters on GAN performance and demonstrates the capability of GANs in creating artificial thermal image datasets.