Leakage-Controlled Multi-View Probability Fusion for DGA-Based Power Transformer Fault Classification: Optimization Ablation and Robustness Analysis

Dissolved gas analysis is universally used to identify power transformer faults; however, direct neural network optimization can be underpowered when highly correlated gas variables are fed to a small classifier. This study evaluates a leakage-controlled multi-view probability-fusion framework and explicitly tests whether Hybrid Particle Whale Optimization Algorithm (HPWOA) initialization is necessary for the final classifier. The original 595-observation IEC 60599-labelled dataset was audited; six exact predictor duplicates carrying contradictory labels were removed, leaving 589 quality-controlled records. A fixed stratified split produced 471 training and 118 held-out observations. Core15 and Extended34 feature views were modelled with six Random Forest, Extra Trees, XGBoost, and LightGBM configurations, and five-fold out-of-fold probabilities formed a 36-dimensional input to a six-class softmax fusion layer. In the original optimizer comparison, canonical HPWOA achieved 76.27% held-out accuracy, whereas HPWOA-initialized ten-iteration L-BFGS refinement reached 82.20% accuracy, 82.47% balanced accuracy, 82.51% Macro F1, and 0.9551 Macro-AUC; however, ordinary logistic stacking produced the same held-out correctness pattern. A controlled initialization analysis further showed that the default start and 30 random starts converged to the same hard-label L-BFGS solution, providing no evidence that HPWOA initialization is necessary for the convex fusion layer. The six-model fusion improved held-out accuracy by only 0.84 percentage points over LGBM-Combined49 (81.36%), while adding substantial ensemble and stochastic-search cost. Repeated 5-fold cross-validation over 10 repetitions for the strongest single learner yielded 75.04% ± 4.41% accuracy, and a similarity-grouped leakage stress test yielded 67.39% ± 8.85% accuracy. Held-out accuracy also declined from 87.50% in the nearest train–test similarity stratum to 71.79% in the most dissimilar stratum. These results indicate that the principal contribution lies in leakage-controlled multi-view probability representation and rigorous optimization/robustness ablation rather than an HPWOA-specific predictive advantage. External and transformer-disjoint validation remain necessary before cross-utility generalization can be claimed.