GXL-TFD: INTELLIGENT DGA-BASED FAULT DIAGNOSIS AND REMAINING USEFUL LIFE PREDICTION FOR POWER TRANSFORMER

Abstract

Internal faults and insulation degradation in power transformers result in costly outages and supply disruptions. This is why we need automated diagnostic systems that can accurately find faults, estimate how much useful life a transformer has left, and send us reports that we can act on. This paper proposes GXL-TFD(Gradient boosting eXplainableAI LLM Transformer Fault Diagnostic), a complete smart DGA framework that includes three types of Gradient Boosting (Gradient Boosting, XGBoost, and LightGBM) for fault detection and diagnosis, a Gradient Boosting Regressor with physics-based synthetic labels for predicting how much useful life something still has, a deterministic Explainable AI module based on built-in feature importance for clear diagnostic reasoning, and a GPT-4.1-nano Large Language Model for automatically creating IEC 60599- compliant diagnostic and prognostic report generation. Using 10-fold stratified cross-validation with SMOTE inside each fold, the GXL-TFD FDD classifier was tested on the MDPI DuVal dataset (447 samples, 3 fault classes). The proposed framework achieved 96.43% accuracy with effective precision, recall, and F1-score, while the RUL regressor got a R² of 0.9935 and an RMSE of 0.58 % on the Mendeley dataset (1,866 samples). GXL-TFD makes it possible for black-box machine learning predictions to become maintenance decisions that can be put into action by giving accurate, hallucination-free diagnostic reports at very low computational cost.