Comprehensive Evaluation of Machine Learning Models for DGA-Based Fault Classification in Power Transformers

Abstract

Dissolved Gas Analysis (DGA) is a widely used technique for condition monitoring and fault diagnosis of oil-immersed power transformers. Conventional DGA interpretation methods, including ratio-based and graphical approaches, are simple to implement and widely adopted in industrial practice. However, these methods may lead to ambiguous or unidentified fault diagnoses under complex operating conditions. To address these limitations, machine learning approaches have recently attracted significant attention for improving DGA-based fault classification. This paper presents a systematic comparative evaluation of several supervised ML models for transformer fault diagnosis using DGA data. The performance of these models is benchmarked against widely used conventional rule-based DGA interpretation techniques, including the Rogers method, the Duval Triangle, and the Duval Pentagon. To provide a comprehensive comparison, the evaluation is conducted under two complementary data regimes: a full dataset that reflects realistic conditions with incomplete measurements, and a reduced dataset containing only samples with complete gas measurements. A unified evaluation framework is adopted to ensure a fair comparison using standard classification metrics, including accuracy, recall, and F1-score, reported in both macro-averaged and weighted forms. The results provide a comprehensive assessment of the strengths and limitations of ML-based and conventional rule-based DGA interpretation techniques and offer practical insights for the application of data-driven methods in transformer condition assessment.