Data-driven Fault Detection and Diagnosis of Power Transformers Using Dissolved Gas Analysis and Random Forest Classifiers

Power transformers are critical and high-value assets in electric power systems, and their unexpected failure can lead to severe economic losses, safety hazards, and prolonged service interruptions. Dissolved Gas Analysis (DGA) is widely used for transformer condition monitoring. However, conventional interpretation techniques rely on fixed thresholds and heuristic rules. These methods often struggle under complex, overlapping, or incipient fault conditions. This study proposes a data-driven framework for transformer fault detection and diagnosis. The framework integrates DGA with a Random Forest classification model. Its purpose is to improve diagnostic reliability and interpretability. Historical, labelled DGA data comprising hydrogen, carbon monoxide, ethylene, and acetylene concentrations were analysed and classified into normal operation, partial discharge, overheating, and arcing fault categories. To enhance model robustness, multicollinearity was mitigated through feature selection, while class imbalance was addressed using the Synthetic Minority Over-sampling Technique. The Random Forest classifier was trained with optimised hyperparameters and evaluated using precision, recall, F1-score, confusion matrix analysis, and out-of-bag error estimation. The results demonstrate high diagnostic accuracy for normal operating conditions and partial discharge faults, with strong precision and recall, while moderate performance was observed for overheating and arcing faults due to inherent overlap in gas generation patterns. Feature importance analysis further revealed the relative contributions of key dissolved gases, enhancing model transparency and engineering insight. The findings confirm that ensemble learning can effectively capture nonlinear relationships in DGA data that are not addressed by conventional methods. This work contributes an interpretable and practical diagnostic framework that supports predictive maintenance, informed decision-making, and improved reliability of transformer condition monitoring in modern power systems.