From limited data to reliable diagnosis: an interpretable deep learning framework for transformer fault analysis

• A GAN-based data augmentation method generates realistic DGA samples, reducing dataset sparsity and imbalance. • Hyperparameters of the CNN-BiLSTM-Attention model optimized using the GOOSE algorithm, improving training efficiency and diagnostic performance. • SHAP analysis reveals feature contributions, enhancing model interpretability and aligning predictions with domain knowledge. The reliable diagnosis of power transformer faults is important for ensuring the safety and stability of modern power systems. However, existing fault identification techniques suffer from limited diagnostic accuracy due to insufficient feature representation, inadequate handling of data imbalance in Dissolved Gas Analysis datasets, and suboptimal model generalization. Furthermore, the absence of comprehensive theoretical investigations into the underlying fault mechanisms and model interpretability has significantly constrained the development of robust and explainable diagnostic frameworks. To address these problems, the GAN-CNN-BiLSTM-Attention-GOOSE framework was proposed to overcome the limitations of traditional transformer fault diagnosis, address data scarcity challenges and provide new avenues for future transformer protection research. In this study, an attention-based deep learning model was developed to improve the accuracy and reliability of transformer fault diagnosis. To figure out the limitations posed by insufficient data, a generative adversarial network was introduced to enrich the training samples. A GOOSE optimization algorithm was employed to fine-tune the learning process and enhance overall performance. This integrated approach yielded higher classification accuracy compared to conventional methods. To further interpret the model’s predictions, an explainability technique was applied to analyze the input gas data. The analysis revealed clear patterns linking specific gas compounds in transformer oil to operational faults. In particular, two key indicators, C 2 H 2 and C 2 H 4 , were found to be strongly associated with high-energy arcing and thermal faults, respectively. These findings highlight the importance of adequate training data and careful model calibration in achieving accurate and interpretable fault identification.