Fault Diagnosis Method for DC Current Transformers Based on Deep Learning Hybrid Model

ABSTRACT To address the challenges of scarce fault samples, significant noise interference, and complex feature coupling in DC current transformer (DCCT) fault diagnosis, this paper proposes a deep learning hybrid model‐based approach. First, a conditional Wasserstein generative adversarial network with gradient penalty (CWGAN‐GP) is employed as a data augmentation module. By introducing gradient penalty to enforce the Lipschitz continuity of the discriminator and incorporating fault‐type labels for conditional control, this module effectively expands the representation diversity of minority fault classes. Next, an improved multi‐head spatiotemporal Transformer encoder is designed to deeply mine multi‐modal signals. It utilizes parallel temporal and spatial attention mechanisms to extract dynamic evolution patterns in current signals and spatial correlations among temperature sensors, respectively, while a cross‐modal interaction layer models long‐range dependencies among current, temperature and vibration multi‐modal signals. Finally, an adaptive graph convolutional network (adaptive‐GCN) is innovatively proposed to fuse topological features of sensor networks. This module dynamically constructs functional association graphs among sensor nodes through a learnable graph structure generation mechanism and achieves cascaded fusion of multi‐scale spatiotemporal features in the propagation layer, significantly enhancing DCCT fault diagnosis performance. Experimental results demonstrate that the proposed model achieves the highest Macro‐F1, accuracy, and MFSI scores on two datasets: 96.58%, 97.13%, and 3.86 (Dataset 1) and 96.32%, 97.22%, and 3.89 (Dataset 2), outperforming several comparative methods. Furthermore, an edge computing‐based ‘cloud training‐edge inference’ hierarchical architecture is designed to optimize the model inference process, and a knowledge distillation‐driven lightweight model adaptation method is proposed. Experimental results in industrial edge scenarios show that the optimized model reduces the end‐to‐end fault diagnosis response time by 80.9% while maintaining high diagnostic accuracy, meeting the millisecond‐level response requirements of engineering field deployment.