Defect Recognition for Vehicle Cable Terminals in EMUs Using Data Augmentation and Improved ResNet Model

Defect diagnosis for vehicle cable terminals (VCTs) is essential for the safe and stable operation of electric multiple units (EMUs). Because online monitoring of partial discharge (PD) in VCT is not feasible, there is a scarcity of field PD data and an imbalanced data distribution. To address this issue, we first compared and analyzed the effectiveness of three typical improved generative adversarial networks (GANs) in generating PD data. Then, we developed an improved ResNet18 model that incorporates convolutional block attention modules (CBAMs) to improve feature extraction and representation capabilities. Finally, we proposed a combined approach that utilizes data augmentation alongside an improved ResNet model to accurately diagnose internal defects in VCT. The results indicate that the deep convolutional GAN (DCGAN) achieves the highest performance regarding image structure, texture, and diversity of generated samples, effectively reducing the cost of on-site PD data acquisition. Furthermore, the improved ResNet18 model outperforms the traditional ResNet18 model across various performance metrics, including accuracy, recall, ${F}1$ -score, and G-mean, on both small sample and augmented datasets, surpassing other prominent deep learning (DL) models such as AlexNet and VGG-19. The proposed combined diagnostic method increases the accuracy of defect recognition for internal defects in VCT from 86.8% to 98.5%, effectively addressing the problems of low accuracy and limited generalization in diagnostic models caused by the scarcity of field PD data.