Transformer Winding Fault Diagnosis Method Based on GAF Images and Improved Swin Transformer Network Models

Transformers are critical equipment in power transmission and distribution systems. The condition of transformer windings critically affects their reliable operation, making fault diagnosis of transformer windings crucial. This paper proposes a transformer winding fault diagnosis method based on the Gramian Angular Field (GAF) and an enhanced Swin-Transformer network model (SwinDCC). First, an experimental platform for simulating transformer winding faults was established. Using Frequency Response Analysis (FRA), characteristic curves were acquired under three fault types (axial displacement, Inter-pancake short circuit, and bulge warping) across three distinct fault regions—laying the data foundation for subsequent intelligent diagnosis. Subsequently, leveraging GAF image conversion technology, frequency response curves are transformed into GADF (Gramian Angular Difference Field) and GASF (Gramian Angular Summation Field) images. The Swin-Transformer network is then augmented with three key enhancements: Convolutional Block Attention Mechanism (CBAM), Context Fusion Module (CFM), and Depthwise Separable Convolution (DSC). Through comparative analysis of diagnostic accuracy across different fault types and regions against baseline models, this work ultimately develops a transformer winding fault diagnosis method integrating GAF with deep neural networks. Finally, the proposed fault diagnosis method was applied to in-service transformers for comprehensive performance validation. The results demonstrate that using GADF and GASF images as input for convolutional neural networks achieved diagnostic accuracy rates exceeding 90% for both fault types and fault regions, confirming the efficacy of GADF and GASF images as input features for CNN. Experimental results indicate superior classification accuracy when utilizing GADF images as the dataset input. Specifically, the GADF-SwinDCC combination achieved the highest diagnostic accuracy, reaching 97.91% for fault types and 95.83% for fault regions. Compared to the baseline Swin-Transformer model with GADF input, the optimized SwinDCC model demonstrated significant accuracy improvements of 6.24% for fault types and 4.58% for fault regions, validating the proposed method's capability for precise diagnosis of diverse transformer winding faults across varying locations.