Abstract: Compared with the traditional methods such as the Three Ratios, the methods for transformer fault diagnosis based on the machine learning algorithms have the diagnostic efficiency and accuracy advantages. However, the inherent attribute of the "black-box model" determines the inexplicability of the decision-making process and the diagnostic results. To solve this problem, an interpretable transformer fault diagnosis based on the dissolved gas analysis of the transformer oil is proposed in this paper. The tree shapely additive explanations (TreeSHAP) method is adopted to realize the interpretability analysis of the fault diagnosis model based on the tree-structured parzen estimator-extreme gradient boosting (TPE-XGBoost). First, a 24-dimensional fault feature set covering the multi-structural data such as the content, the ratio, and the code of the dissolved gas in the oil is constructed, and 10 available features are picked out and obtained. Then, the transformer fault diagnosis model based on the TPE-XGBoost is proposed. The tree-structured parzen estimator is used to complete the multi-parameter synchronous optimization of the XGBoost model, realizing the accurate judgment of the fault types. Finally, the TreeSHAP theory is introduced to analyze the interpretability of the transformer fault diagnosis model, achieving the visualization of the fault diagnostic decision-making process and its influencing factors and obtaining the key features of different fault types. The research shows that the average accuracy of the transformer fault diagnosis proposed in this paper reaches 90.23%, and the influence process and degree of the features on the model decision-making can be reflected. This method has good accuracy, robustness, and interpretability, which will provide targeted guidance and suggestions for the transformer operation and maintenance.