A New Data Augmentation Model for Fault Diagnosis of Transformer Windings Under Scarce Fault Data

ABSTRACT Deep learning methods for fault diagnosis in transformer windings have drawn growing research interest in recent years. However, the scarcity of FRA fault data limits the application of deep learning in this field. To address this challenge, we propose a data augmentation model based on the diffusion model, namely, CDFF‐TW, to generate enough FRA fault data used in the field. First, this study designs a new input data format in the proposed model, which replaces traditional data preprocessing with the direct input of data. Moreover, the forward diffusion process of the proposed model is designed to be reusable, greatly reducing the overall training time compared to traditional diffusion models. Subsequently, a classifier is incorporated into the reverse diffusion process, and a screening component is designed to screen generated data based on the morphological similarity and the resonance peak similarity. Furthermore, a series of comparative tests is conducted. In addition, the model used the CDFF‐TW to augment data has better performance compared to existing models in transformer winding fault diagnosis. Finally, the optimal ratio of generated to original data is provided as a guideline for augmenting data in fault diagnosis models using the proposed method.