Accurate transformer fault diagnosis based on dissolved gas analysis is essential for reliable power system operation.However, data-driven models often fail to capture the geometric and hierarchical structure of the Duval triangle because they mainly learn statistical correlations without explicit physical constraints.This limitation results in unstable predictions near class boundaries and under imbalanced conditions.In this study, we present a Duval-guided learning framework that embeds diagnostic knowledge directly into the training objective.The proposed framework includes a Duval soft alignment loss based on KL divergence to preserve geometric relationships, an ordinal constraint to enforce energy level progression within fault groups, and a gating mechanism to maintain hierarchical consistency between thermal and discharge faults.These components guide the model toward physically meaningful and diagnostically consistent probability distributions.Experimental results show that the proposed approach achieves 85.12% accuracy, indicating that explicit integration of domain knowledge improves robustness and reliability in transformer fault diagnosis.
An et al. (Tue,) studied this question.