To enhance the accuracy of transformer fault diagnosis and mitigate the impact of sample imbalance on model recognition accuracy, a transformer fault diagnosis method based on SMOTE and GWO-XGBoost is proposed. This method utilizes SMOTE to expand minority samples, employs non-coding ratio methods to construct multidimensional feature parameters, and introduces neighborhood rough set optimization strategies alongside gray wolf optimization of XGBoost parameters. Experimental validation demonstrates that the method significantly reduces misclassification of minority samples, achieving high precision, low misjudgment rates, and stability, making it suitable for practical applications in transformer fault diagnosis.