Winding fault detection is crucial to ensure the safe and reliable operation of transformers. In order to break through the limitations of the traditional frequency response information characterization and to improve the fault diagnosis accuracy

this paper proposes an image fusion framework based on GAF-VSM-WLSO and combined with a novel HDC-CBAM-ResNet34 classification model to realize the accurate diagnosis of transformer winding faults. First

the frequency response information is transformed into two-dimensional GASF and GADF images by Gram angle field

and then the two types of images are fused using visual saliency map with weighted least squares optimization algorithm to generate VSM-WLSO fused images that retain global structure and local details

which solves the problem of low redundancy in the representation of a single feature. Second

the HDC-CBAM-ResNet34 classification model fusing hybrid dilation convolution and attention mechanism is constructed

which takes ResNet34 as the benchmark

introduces hybrid dilation convolution (HDC) so as to expand the receptive field for capturing multi-scale features

and incorporates a convolutional block attention mechanism (CBAM) module to enhance key feature information. Consequently

high-accuracy recognition of the type

degree

and location of transformer winding faults is achieved. Finally

the performance of the model proposed in this paper is verified and comparatively analyzed through ablation experiments and comparisons with current mainstream models. The experiment results show that the method proposed in this paper performs well in the classification of fault type

fault location

and fault degree

and its recognition accuracy and F1 parameter are above 96%.