Abstract: A transformer fault diagnostic method based on Borderline-SMOTE-CatBoost-SHAP is proposed for the problem of degradation of classification performance of fault diagnostic model and uninterpretable model diagnostic results caused by unbalanced distribution of transformer fault data. Firstly, the Borderline-SMOTE algorithm is used to equalize the data while retaining the distribution characteristics of a few classes of samples, solving the bias problem caused by the imbalance of fault data distribution. And then, the CatBoost transformer fault diagnosis model is constructed, and simulation experiments are carried out by us‍ing the actual fault data of the transformer to compare it with different data equalization methods and other transformer fault diagnosis models. Finally, the SHAP model is introduced to provide a more accurate and reliable classification model for transformer fault diagnosis. The SHAP model is introduced to analyze the interpretability of the CatBoost fault diagnosis model to solve the problem of poor interpretability of the “black box model”. The result shows that the overall diagnostic accuracy of the transformer fault diagnosis model based on Borderline-SMOTE-CatBoost-SHAP is 92.99%, the F₁ value is 0.91, and the Kappa coef‍ficient is 0.91. Meanwhile, the process of input features for decision-making is visualized to verify the validity of the proposed method.

Key words: unbalanced samples, CatBoost, fault diagnosis, shapely additive explanation, interpretability