Transformers are core equipment in power systems, and their failure may cause severe accidents. Dissolved gas analysis (DGA) is one of the dominant techniques for fault diagnosis of oil-immersed transformers. To achieve lightweight design and high identification efficiency, this paper proposes an early fault identification method for transformers based on the integration of the Convolutional Block Attention Module-enhanced ShuffleNetV2 (CBAM-SV2) model and Gramian Angular Field (GAF). First, hybrid oversampling is used for data preprocessing. Then, the preprocessed one-dimensional gas data are converted into dual-channel two-dimensional images via GAF as the input of the classification network. Finally, a CBAM-SV2 model integrating deep convolutional networks and attention mechanisms is constructed, which combines the lightweight advantage of ShuffleNetV2 and the powerful feature representation ability of the Convolutional Block Attention Module (CBAM). Feature extraction and classification are performed by the CBAM-SV2 model to output the identification results. Additionally, t-distributed Stochastic Neighbor Embedding (t-SNE) and a confusion matrix are used to visualize classification performance for intuitive evaluation of the network’s effectiveness. The experimental results show that, compared with other mainstream algorithms, the proposed method achieves higher recognition accuracy in transformer early fault classification under imbalanced data conditions.
Yang et al. (Fri,) studied this question.