To address the insufficient utilization of feature information from easily confusable samples in existing transformer fault diagnosis models

this paper presents a transformer fault diagnosis model based on Shapley additive explanations (SHAP) primary-secondary feature selection and heterogeneous dual-branch basis spline-wavelet Kolmogorov- Arnold network (B Spline-WavKAN). First

the SHAP algorithm is used to quantify the contribution of each feature to different fault types

achieving the effective differentiation between primary and secondary features. Second

a B Spline-WavKAN heterogeneous dual-branch feature extraction and adaptive fusion framework is introduced. Primary features are fed into the B SplineKAN to learn global features

while secondary features are fed into the WavKAN to deeply extract boundary information of easily confusable samples. The primary and secondary features are dynamically weighted and complemented through a confidence-guided adaptive fusion mechanism. Finally

the fused enhanced features are fed into a lightweight gradient boosting machine (LightGBM) for fault diagnosis. Experimental results demonstrate that the proposed method can efficiently capture fault feature information and improve the recognition of easily confusable samples

achieving a diagnostic accuracy of up to 98.7% and demonstrating strong potential for engineering application.