Publication

Abstract : Fault identification in transformers is a challenge in power systems, as data collected across different years often exhibits domain shifts, resulting in significant declines in model performance. Classic deep learning models perform well on the source domain but struggle to generalize when a change occurs in the target domain distribution. To address this, a unified approach is put forward that integrates deep neural feature extraction with cutting-edge domain adaptation techniques. Domain-invariant representations are statistically aligned using Maximum Mean Discrepancy (MMD) and further refined through contrastive feature alignment. Moreover, adversarial training with a Gradient Reversal Layer (GRL) is applied to reinforce true domain invariance and improve cross-domain robustness. Entropy minimization, pseudo-labeling, and consistency regularization also enhance target domain predictions in weak supervision settings. The hybrid learning process ensures high source domain accuracy while also adapting to new domains simultaneously.The proposed method is evaluated on multi-year transformer datasets and achieves superior performance over state-of-the-art approaches.