Acoustic signal augmentation for fault diagnosis of power transformers based on improved cycle generative adversarial networks

Modeling and analyzing the voiceprint properties of operational acoustic signals from power transformers has emerged as a promising and effective method for fault diagnosis. However, acquiring sufficient fault data for training diagnostic models presents a challenge due to the rarity of fault operational events in practical applications. To address this problem, this study proposes an acoustic signal augmentation method for fault diagnosis of power transformers based on improved cycle generative adversarial networks (CycleGAN). First, the generator and discriminator in the basic CycleGAN model are optimized and a comprehensive objective function is proposed for improving the model's ability to capture and learn the non-linear features of acoustic signals effectively. Second, an acoustic signal augmentation model, capable of generating acoustic signals under six fault statuses, is proposed based on the improved CycleGAN to effectively augment the limited training dataset of practical power transformers, thereby enhancing the availability of diverse training samples for fault diagnosis. Finally, a fault diagnosis system that integrates our signal augmentation model is constructed, and three conventional classifiers are respectively combined to comprehensively assess our model's effectiveness and practicability. The experimental results demonstrate that the classification performance of existing fault diagnosis methods can be significantly improved with our augmented training dataset, especially the classification accuracy is improved by nearly 20%.