Enhanced Fault Diagnosis for New Energy Vehicle Systems Using Improved Places365GoogLeNet with Feature Pyramid Network and Swin Transformer

This study aims to solve the problem of insufficient accuracy in fault detection and diagnosis of the three-electric system of new energy vehicles (NEVs). By improving the Places365GoogLeNet model, combined with the feature pyramid network (FPN) and Swin Transformer structure, the multi-scale feature modeling capability is improved to achieve efficient and reliable fault prediction and diagnosis. The operation information of NEVs is collected, and after preprocessing the data, the continuous fault records are converted into image data using a sliding window method to construct a data set. In terms of model design, this paper introduces FPN based on the traditional Places365GoogLeNet, which has the ability to fuse multi-scale features and process multi-level information in complex scenes. By embedding Swin Transformer, the model’s modeling capabilities for time series data and spatial features are enhanced. The accuracy of the model in this paper is 96.2%, and the macro precision, macro recall and macro F1 are 97.2%, 97.1% and 97.1%, respectively, which are significantly better than ResNet50 (Residual Networks 50) (accuracy 95.3%) and ResNet101 (accuracy 95.8%). In addition, the training time of the model in this paper is only 7,102[Formula: see text]s, which is highly time-efficient. The ablation experiment further verifies the independent contributions of FPN and Swin Transformer. The improved Places365GoogLeNet model combining FPN and Swin Transformer performs well in NEV fault classification, providing an efficient and accurate solution for intelligent fault prediction and vehicle health management.