Interpretable analysis of transformer winding vibration characteristics: SHAP and multi-classification feature optimization

Abstract

Current research on vibration-based winding looseness struggles to clearly determine the contribution of individual features to the model. This paper proposes using the distribution of the harmonic-to-fundamental frequency ratio as a feature for a 110 kV transformer and introduces the SHapley Additive exPlanations (SHAP) method to analyze feature contributions. However, since SHAP tends to overlook the recall of certain states in multi-classification problems, a feature subset optimization scheme is proposed. After conducting cross-validation with multiple classifiers and analyzing the impact of combined features, SHAP analysis is performed on the constructed features to generate a key feature union set. Once thresholds for overall accuracy and sub-state recall are set, features are sequentially removed from the union set while their compliance with the predefined thresholds is evaluated, thereby determining their retention. Experimental results show that the proposed method achieves 99.73 % accuracy in identifying winding looseness states, improving by 0.03 % compared to SHAP, while enhancing the recall of the severe fault state A3 by 0.26 %. Overall, the proposed method effectively balances accuracy and key state recall, offering a new perspective on feature analysis in fault diagnosis.