Adaptive Rogers’ Ratio-Based Transformer Fault Diagnosis: A New Hybrid Fungal Growth Optimizer with Teaching-Learing-Based Optimization (FGO–TLBO) Metaheuristic for Explainable Intelligent Optimization Framework

Abstract

Accurate identification of transformer insulation faults is essential for reliable power-system operation, and dissolved gas analysis (DGA) remains an important diagnostic tool. However, conventional Rogers’ Ratio interpretation relies on fixed decision boundaries that may not represent overlapping gas patterns and measurement variability. This study proposes an explainable Hybrid FGO–TLBO framework that adaptively calibrates four Rogers’ Ratio decision boundaries while preserving the original rule-based diagnostic logic. The study uses 500 real DGA records representing six classes: Normal, Partial Discharge (PD), Low-Energy Discharge (D1), High-Energy Discharge (D2), Low-Temperature Thermal Fault (T1), and Medium-Temperature Thermal Fault (T2). A stratified 80/20 split was used, with 400 records for threshold calibration and 100 untouched records for held-out evaluation. The optimization employed 50 independent runs, 100 iterations, a population size of 50, 2% Gaussian perturbation during calibration, and a penalty coefficient of λ = 0.05. The proposed Hybrid FGO–TLBO achieved a mean calibration fitness of 0.70892 and a standard deviation of 2.41 × 10−4 over 50 runs. On the held-out test set, the adaptive framework achieved 83% accuracy and a macro-F1 score of 0.7701, compared with 67% and 0.6064, respectively, for conventional Rogers’ Ratio. The optimized thresholds were x1 = 0.227385, x2 = 1.337761, x3 = 0.147130, and x4 = 0.556188. The results support hybrid metaheuristic calibration as an interpretable threshold-optimization approach, while the held-out results are presented as an internal validation and not as evidence of universal superiority or guaranteed optimality.