Power transformer health index assessment: An explainable LightGBM based minority aware learning using SMOTE-ENN hybrid resampling

Power transformers are an expensive and critical component of every power system. Reliable assessment of transformer condition evaluation is the backbone of preventive maintenance strategies. Transformer condition majorly relies on its oil insulation and effective assessment of insulation enables early fault identification, transformer condition assessment and reliability enhancement by collecting testing data without impairing equipment integrity or interrupting operation. These datasets often exhibit severe class imbalance, where certain health states are highly underrepresented. This class imbalance problem leads to biased decisions of conventional machine learning models toward dominant classes. Also, causing a reduction in their capability to detect minority health conditions at an early stage. To overcome this problem, our study introduces an explainable minority-aware framework for power transformer health index classification using LightGBM combined with a hybrid SMOTE-ENN resampling technique. Six resampling methods were further investigated including Random Oversampling, SMOTE, Borderline-SMOTE, K-Means-SMOTE, CTGAN, and SMOTE-ENN to validate their effectiveness in improving minority class learning. The datasets generated via each resampling technique are assessed through application of multiple classification models, and their performance is evaluated using classwise precision, recall, F1-score and overall accuracy to effectively measure minority class behavior. Experimental results validate that hybrid resampling significantly improves minority class detection by balancing the original imbalanced dataset and removing noisy samples as compared to traditional oversampling methods. LightGBM classifier applied to SMOTE-ENN generated dataset achieves best performance, reaching 99% accuracy with balanced classification across all health states, validating the effectiveness of noise aware resampling with gradient boosting in imbalanced transformer datasets. To improve transparency, explainability analysis is then performed on best chosen LightGBM model which identify key influential features and provide clear insights into minority class detection patterns. This proposed framework collectively offers a robust and interpretable solution for health index evaluation, supporting improved maintenance planning and asset management decisions.