Abstract: To solve the problems of existing intelligent fault diagnosis methods such as single feature input, difficulty in extracting faults, and poor model interpretability, an interpretable fault diagnosis method based on XGBoost (extreme gradient boosting) and SHAP (SHapley Additive exPlanations) analysis is proposed. Firstly, traditional signal processing methods are used to complete the extraction of multi-domain features. Secondly, a fault diagnosis model is constructed based on the XGBoost integrated algorithm, and a preliminary feature explanation of the model is performed based on the XGBoost embedded evaluation indicators. Finally, the Tree SHAP method is used to interpret and analyze the features of the diagnostic model, explore the influence of important features on the trend of bearing fault categories, analyze the dependence and interaction effects between features, and intuitively and transparently reveal the diagnostic mechanism of the model. Through experiments comparing XGBoost with other traditional machine learning methods, this model has more outstanding comprehensive performance in multi-dimensional evaluation indicators and has strong accuracy. The fault diagnosis accuracy rate is as high as 99.62%, which has good practical application value.