• A machine learning framework is proposed for non-invasive prediction of transformer paper degradation • Gradient Boosting achieved the highest accuracy for DP prediction, while Random Forest showed strong performance in insulation life estimation • Pearson correlation analysis and SHAP interpretability identified 2FAL and 5HMF as the most influential degradation indicators • Bootstrapping provided the most robust generalization performance among the evaluated resampling techniques The degradation of cellulose-based insulating paper is a leading failure mechanism in power transformers, directly impacting equipment operational safety. Traditional diagnostic methods rely heavily on single chemical markers, like 2FAL, which often limit predictive accuracy. This failure detection approach presents notable limitations, as it neglects other potentially informative furan derivatives and dissolved gases that could enhance predictive accuracy. In this study, a robust machine learning-based framework to predict the degree of polymerization (DP), which is a key indicator of paper insulation degradation and estimate the loss of life (LOL) of transformer paper insulation is developed. By incorporating all the furan derivatives and the CO 2 /CO gas ratio as degradation markers and utilizing ensemble learning models enhanced with Pearson correlation analysis and Shapley Additive Explanations, the framework improves the reliability of early failure prediction. Gradient boosting regression achieved superior performance, attaining a correlation coefficient (R 2 ) of 0.9838 for DP prediction and 0.9952 for loss of life estimation. These results highlight the potential of data-driven predictive maintenance strategies in preventing catastrophic transformer failures, extending asset lifespan, and supporting reliability-centered asset management.
Andrew Adewunmi Adekunle (Tue,) studied this question.