Aiming at the problem of low accuracy of transformer fault diagnosis, a random forest-recursive feature elimination (RFRFE) algorithm and an improved sparrow algorithm ( ISSA) optimization of the extreme gradient boosting tree (XGBoost) transformer fault diagnosis method are proposed. First, based on the diagnostic accuracy, the RFRFE algorithm is used to select important feature variables to remove redundant features; then the traditional sparrow algorithm ( SSA) is improved by the uniform distribution random adjustment strategy and the Levi flight strategy, and the ISSA and SSA and particle swarm optimization ( PSO) performs algorithm performance testing, which proves that its classification accuracy and network optimization capabilities have been improved; finally, the improved sparrow algorithm is used to optimize XGBoost related hyperparameters to obtain the synthesis of the combination of RFRFE and ISSA-XGBoost. The fault diagnosis model is compared with the PSO-XGBoost and SSA-XGBoost fault diagnosis models. The results show that the fault diagnosis rate of ISSA-XGBoost is 91. 08%, which is 9. 9% and 6. 93% higher than that of PSO-XGBoost and SSAXGBoost, respectively. The proposed method can effectively improve the performance of transformer fault diagnosis.