Deep learning methods have shown great potential in the field of fault diagnosis of train wheelset bearings， but their effec? tive implementation is based on the correct matching between various types of data and category labels. For data with a small num? ber of label error samples， traditional deep learning methods are difficult to achieve the expected diagnostic effect. To address this issue， this paper proposes a fault diagnosis method combining box graph method and feature fusion model is proposed to address this issue. In this method， the outlier in each group of bearing signals is removed by box graph method， and the remaining data is expanded by the SMOTE method to restore to the original data size； Input the processed sample data into the improved feature fu? sion model for fault identification and classification. The experimental data of train wheel bearings was used for validation. The re? sults showed that compared to directly using traditional neural network models for fault diagnosis， the diagnostic accuracy of the method proposed in this paper is higher， indicating that the method has better processing performance for bearing data with a small number of label error samples.