Transformer fault diagnosis method based on Gramian Angular Field and optimized parallel ShuffleNetV2.

To address the issue of low diagnostic accuracy caused by reliance on manual experience and single-angle feature extraction in transformer fault diagnosis, a transformer fault diagnosis method based on Gramian Angular Field and optimized parallel ShuffleNetV2 was proposed. First, the fault signals were mapped into two types of feature images, GASF and GADF, constructing an image sample set that fully reflected fault information. Then, an optimized dual-branch parallel ShuffleNetV2 model was built to simultaneously perform efficient feature extraction on the two types of images, and the Convolutional Block Attention Module was introduced to achieve adaptive weighted fusion of features. Finally, the fused features were input into the SoftMax classifier for classification and identification, achieving transformer fault diagnosis. Experimental results showed that the propsed model achieved an overall accuracy of 99.13%, demonstrating strong robustness and generalization performance. The proposed method not only avoided the limitations of manually extracted features, but also achieved a large improvement in diagnostic effectiveness compared to single-angle feature images.