KGRAT: An IEC-Informed Knowledge Graph Attention Representation for Power Transformer DGA Diagnosis

Le résumé fourni par la source

Dissolved gas analysis (DGA) is widely used for power transformer fault diagnosis, but many learning-based studies still treat gas concentrations and derived ratios as flat input features. KGRAT is positioned here as an IEC-informed graph representation with a relation-conditioned graph attention learner, rather than as a universally strong predictor. Gas, symptom, and fault entities are linked by four standards-informed relation types, and relation-conditioned attention is learned over this fixed graph. On a six-class benchmark of 589 samples evaluated with stratified 10-fold cross-validation, KGRAT achieved 0.7233 accuracy and 0.7089 Macro-F1. In this single-seed evaluation, it scored above IEC Three-Ratio, Duval Triangle, raw-feature SVM, raw-feature MLP, and a complete-graph GAT ablation; the dependent-fold Holm diagnostic supported the complete-graph contrast within that run but is not seed-robust inference. Feature-engineered tree ensembles were stronger, with GBDT using ratio/symptom features reaching 0.8283 Macro-F1. A filtered four-label Cliango/DGA evaluation is reported only as a constrained stress test over common labels, not as six-class external validation. The evidence therefore supports KGRAT as a standards-aligned, relation-level inspectable representation for DGA modeling, not as a deployment-ready diagnostic system or a substitute for stronger feature-engineered tree ensembles on this dataset.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.