A Novel Transformer Fault Diagnosis Method for Imbalanced Data Based on Dual Attention Mechanism and IEC Ratio Features

Accurate and rapid fault diagnosis of transformers is crucial to ensuring the safety and stability of power systems. However, the scarcity and imbalance of fault data lead to inadequate adaptability of existing methods when handling complex patterns, resulting in low diagnostic accuracy. To address this challenge, this paper proposes a deep learning model based on a dual attention mechanism and IEC ratio features (DAM-IRF) for transformer fault diagnosis with imbalanced data. The proposed model combines a Transformer encoder with Convolutional Neural Networks (CNN), incorporating both channel and spatial attention mechanisms to extract global and local features effectively. Additionally, IEC ratio features are encoded and integrated into the learning process, enabling the model to incorporate rule-based diagnostic knowledge alongside data-driven learning. Extensive experiments on dissolved gas analysis (DGA) datasets with varying degrees of imbalance demonstrate the robustness and effectiveness of the proposed method. Specifically, the DAM-IRF model achieves fault classification accuracies of 95.56%, 93.52%, 92.67%, 89.33%, and 87.62% across five increasingly imbalanced cases, respectively, with corresponding F1-scores above 0.87. These results confirm the superiority of the proposed method in handling data imbalance and its promising potential in real-world transformer fault diagnosis applications.