Cyber-Resilient Dissolved Gas Analysis Fault Diagnosis for Power Transformers Using Physics-Constrained Semi-Supervised Adversarial Learning

Dissolved Gas Analysis (DGA) is the primary incipient-fault diagnostic technique for oil-immersed power transformers. Field datasets suffer from label scarcity, severe class imbalance, and growing cybersecurity exposure. This paper proposes MRC-SSAT, a unified physicsgrounded framework that simultaneously addresses all three challenges. A majority-vote pseudo-labeling strategy fusing IEC 60599, the Duval Triangle, and the Duval Pentagon generates supervisory signals from unlabeled DGA data with strong inter-standard agreement (κ = 0.76–0.84), independently validated at 91.3% accuracy against expert-verified IEC TC10 labels. A DGA-specific physics-constrained conditional tabular GAN (sTableGAN) resolves class imbalance by enforcing IEC 60599 gas-ratio plausibility during synthesis. A semi-supervised Transformer encoder with Virtual Adversarial Training (VAT) and domain-aware perturbations ensures decision-boundary smoothness and cyber resilience. SHAP-based explainability confirms that learned gas importances align with IEC 60599 electrochemical principles. Benchmarking on the Lewis Transformer B field dataset confirms statistically significant superiority over thirteen state-of-the-art methods: 97.1%±0.6% accuracy (MCC= 0.94) with only 5.3 pp accuracy degradation under projected gradient descent (PGD) attacks. Cross-dataset validation on IEC TC10 (95.8%) and CIGRE A2.43 (96.3%) confirms generalization across diverse transformer populations. Critically, Lewis B carries no expert labels; IEC TC10 and CIGRE A2.43 results therefore constitute the primary generalization evidence.