Physics-Informed Guided Transformer Fault Diagnosis Via DGA Data

In response to the challenges of limited transformer fault data samples and insufficient fault characteristic representation, resulting in limited diagnostic accuracy. This study proposes a physics-informed guided transformer fault diagnosis (PIGTFD) method to achieve synergistic optimization of feature learning and fault diagnosis processes. Firstly, a feature extraction method based on multi-level contrast-enhanced autoencoder is developed to enable contrast-enhanced learning and multi-level feature refinement, thereby enhancing the characterization of fault data features. Secondly, a physics-informed guided classifier incorporating comparative learning loss, L₂/L₄ norm, and cross-entropy loss is constructed to collaboratively optimize feature extraction and classification. Additionally, a δ-Batch Normalization method is designed to enhance the model’s generalization robustness under data distribution shifts. The model’s validity is confirmed using dissolved gas data from a 110kV power transformer at a substation in Jiangsu Province. Experimental results demonstrate that PIGTFD outperforms other comparative methods, and the accuracy of fault identification is 91.05%.