FAULT DIAGNOSIS IN POWER TRANSFORMERS USING RESAMPLED DISSOLVED GAS ANALYSIS AND MACHINE LEARNING

Fault diagnosis in power transformers is essential for ensuring operational reliability and minimizing maintenance downtime. This paper presents a hybrid machine learning framework that leverages advanced resampling and optimization techniques to enhance fault classification accuracy using dissolved gas analysis (DGA). To address class imbalance in the dataset, resampling methods such as SMOTE-ENN and ADASYN are employed. Three classifiers, namely nonlinear support vector machine (NLSVM), linear discriminant analysis (LDA), and artificial neural networks (ANN) are evaluated, with hyperparameter tuning performed using gradient-based optimizer function (GFO) and firefly optimization (FFO). Among them, the FFO-optimized NLSVM combined with SMOTE-ENN achieved the highest performance, with a recognition rate of 98.7%. The results, validated through confusion matrices and ROC analysis, demonstrate the robustness of the proposed approach. This framework provides an effective and reliable tool for condition-based maintenance, enabling precise multi-class fault identification in power transformers.