A Transformer Fault Diagnosis Method Based on Data Augmentation and IEVO-CNN

Deep learning algorithms have become a trending approach for transformer fault diagnosis. However, challenges such as data imbalance leading to poor model learning performance and difficulties in determining network parameters still exist. This article proposes a transformer fault diagnosis method based on sample augmentation and an improved energy valley optimization (IEVO) algorithm to optimize the parameters of convolutional neural networks (CNNs). First, a combination model of generative adversarial networks and variational autoencoders (GAN-VAE) is established for data augmentation. Then, the energy valley optimizer (EVO) algorithm is improved by integrating Gaussian mixture initialization, dynamic boundary enrichment, and differential evolution strategies. Finally, random forest (RF) is employed to select the optimal 24 candidate feature ratios, and CNN parameters are further refined, resulting in the development of an IEVO-CNN-based transformer fault diagnosis model. Experimental results show that the samples generated by GAN-VAE have an MSE of 0.017 and a KL divergence of 0.186. The IEVO-CNN model achieves a diagnostic accuracy of 91.63%, demonstrating the superiority of the proposed method in handling imbalanced data and enhancing fault classification performance.