Mahmoodiyan, Hootan2025-09-162025-09-162025https://hdl.handle.net/1828/22766Power transformer failures can lead to severe service interruptions and economic loss, making early and accurate fault diagnosis crucial for reliable power grid operation. Dissolved Gas Analysis (DGA) has long been recognized as a standard diagnostic technique; however, the diagnostic performance of machine learning models often degrades when applied to new datasets collected under different operational or environmental conditions—a challenge known as domain shift. This thesis addresses this issue by proposing a robust and interpretable domain adaptation framework tailored to power transformer fault diagnosis. The proposed method, termed MCW (Maximum Mean Discrepancy and CORrelation ALignment with feature-specific Weighting), introduces a novel approach for emphasizing features that exhibit strong statistical differences between source and target domains. Specifically, Kolmogorov--Smirnov (K-S) statistics are employed to compute feature-wise distributional discrepancies, which are then used to weight the contributions of individual features during domain alignment. The hybrid diagnostic features—comprising both conventional and newly derived gas ratios—are transformed into two-dimensional Gramian Angular Field (GAF) images, enabling spatial representation of fault patterns. A custom convolutional neural network (CNN) is trained to classify these images into five fault types. To evaluate the method, experiments are conducted using a source dataset from literature (Egyptian and Indian utilities) and a target dataset from the IEC TC 10 database. The proposed MCW method is compared against baseline Fine-Tuning and conventional MMD-CORAL (MC) approaches using multiple metrics including accuracy, F1-score, Average Kullback--Leibler Divergence (AKLD), and confusion matrices. The results demonstrate that MCW consistently outperforms baseline methods—achieving an average accuracy of 93.6% and F1-score of 93.5%, with notable robustness even under limited labeled target data. Confusion matrices show reduced inter-class misclassifications, and ablation studies confirm the effectiveness of the selected hyperparameters and architecture. Overall, this research demonstrates that incorporating feature-weighted domain adaptation into transformer fault diagnosis pipelines significantly improves diagnostic accuracy and generalization. The findings have practical implications for developing intelligent monitoring systems that remain reliable across diverse grid environments and transformer populations.enAvailable to the World Wide WebPower transformerDomain adaptationTransfer learningMachine learningCNNThesis