Answers are grounded in this paper and cite their sources.
Dissolved Gas Analysis (DGA) is widely used for diagnosing power transformer faults, but performance is constrained by feature redundancy, nonlinear gas interactions, class imbalance, and overlapping fault signatures. Objective: This study develops a multi-view stacked machine learning model for classifying six IEC 60599 fault categories and evaluates whether integrating complementary DGA representations improves performance over ANN and feature-extraction models. Methods: The original dataset contained 595 observations representing partial discharge, low-energy discharge, high-energy discharge, and thermal faults T1, T2, and T3. A data-integrity audit identified six observations with identical gas measurements but contradictory labels; these were excluded, yielding 589 quality-controlled samples. A stratified 80:20 split produced 471 training and 118 testing observations. Core 15-dimensional and extended DGA feature views were modelled using Random Forest and Extra Trees classifiers. Their five-fold out-of-fold class probabilities were integrated through logistic-regression stacking. Raw-gas ANN, IEC three-ratio ANN, unreduced ANN, PCA+ANN, AE+ANN, individual tree learners, and soft voting served as comparators. Results: The proposed model correctly classified 97 of 118 observations, achieving 82.20% accuracy, 82.31% weighted F1-score, 82.57% balanced accuracy, 0.9563 macro-AUC, 0.7848 Cohen’s kappa, and 0.7878 Matthews correlation coefficient. Accuracy improved by 11.86 percentage points over the unreduced ANN and by 2.54 points over the strongest individual learner. The 2.54 pp improvement over the best individual tree models was descriptive and was not confirmed as statistically significant by McNemar testing (p = 0.125 to 0.508). The bootstrap 95% confidence interval was 74.58–88.98%. Conclusions: Multi-view stacking provides statistically confirmed improvements over ANN-based comparators. The 2.54 pp descriptive advantage over the best individual tree models was not statistically confirmed. Generalization claims are observation-level; transformer-disjoint external validation is required before operational deployment.
No takes yet. Share an insight, caveat, or question.
Nkwambe et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: