Fault Diagnosis of Power Transformer Based on Centroid Distance Feature and Explainable Machine Learning

Dissolved Gas Analysis (DGA) serves as the technological foundation for the primary diagnostic method used to detect incipient faults in transformers, with traditional approaches such as the Duval pentagon, Duval triangle, and IEC ratio methods remaining widely adopted in practice. These approaches are fundamentally limited by their deterministic classification thresholds and compromised diagnostic precision. Therefore, this study proposes a transformer defect diagnosis approach utilising traditional DGA feature similarity analysis. First, the collected transformer fault data are mapped in Duval pentagon, Duval triangle and IEC ratio method‐based three‐dimensional coordinates. Ordering Points to identify the clustering structure (OPTICS) is employed to identify fault clusters corresponding to distinct fault types and compute their respective centroids. Second, this study derives a novel discriminative feature based on the Euclidean distance between the acquired DGA data and each fault cluster centroid. The derived feature combination serves as input features for the eXtreme Gradient Boosting (XGBoost) model. Verification shows that the proposed transformer fault diagnosis method achieves a mean classification accuracy of 92.8%. Finally, SHapley Additive exPlanation (SHAP) is employed to evaluate the contribution of different feature indicators to the model results. The proposed model is validated through a field case study, demonstrating both feasibility and explainability.