Traditional transformer health condition assessment primarily relies on industry guidelines and expert experience

typically employing periodic offline evaluations

which makes it difficult to reflect the real-time status of the equipment. Data-driven assessment models

while suitable for continuous tracking of equipment operation and its development trends

face issues such as high demands on raw sample quality and insufficient interpretability. Hence

this paper proposes an interpretable transformer health condition early warning method based on imbalanced data. First

an adaptive synthetic oversampling method is employed to effectively augment the minority class samples

generating a balanced dataset. Subsequently

a transformer health condition early warning model based on Bayesian optimization and lightweight gradient boosting is constructed to achieve precise and efficient forecasting of the transformer's health status. Finally

the Shapley value additive explanation attribution theory is introduced to conduct an analysis of the factors influencing the early warning of transformer health status from both global and individual perspectives

effectively quantifying the impact of each state parameter on the model's predictive outcomes. The research indicates that the proposed method achieves an average accuracy rate of 98.46% in identifying the health status of transformers

effectively reflecting the dynamic interplay between transformer characteristic parameters and the model's predictive results. The results can provide effective support for the intelligent maintenance and the formulation of differentiated maintenance strategies for transformers in operation.