Transformer fault diagnosis is crucial for ensuring the safe and stable operation of power systems. However, traditional diagnostic methods suffer from issues such as insufficient feature utilization and poor model interpretability. To address these challenges, this paper proposes a hybrid fault diagnosis model. First, the model utilizes SHapley Additive exPlanations values for feature selection. Then, the improved Harris Hawks optimization is proposed to optimize the hyperparameters of eXtreme gradient boosting model (Xgboost), further improving the model’s classification performance. Experimental results demonstrate that the proposed model outperforms existing methods across multiple performance indicators, achieving an accuracy of 0.9509, which represents a 2.68% improvement compared to the Xgboost model. This model fully leverages the advantages of each method, effectively solving problems like insufficient feature utilization, poor model interpretability, and limited performance of optimization algorithms in traditional fault diagnosis methods. It provides a more reliable and efficient solution for transformer fault diagnosis.