Diagnosis for transformer winding fault via deep fusion of 2-D images of FRA and SwinLCGnet framework

Winding fault detection is essential to ensure the safe and reliable operation of power transformers. This paper conducts a diagnostic study on winding fault type identification, degree assessment, and spatial localization using frequency response analysis (FRA) test results, a widely used engineering test method. Firstly, we propose the Gramian angular field with frequency-proximity guidance (GAF f ) method, which converts 1-D FRA curves into 2-D images by comprehensively exploiting global and local signal features, thereby achieving improved visualization of winding fault characteristics. Thereafter, the task-customized mixture of adapters (TC-MoA) is applied to fuse GASF f and GADF f feature images for the first time, effectively combining the advantages of dual coding methods to enhance multimodal complementary features. Finally, a two-branch diagnostic framework named SwinLCGnet is proposed to conduct collaborative extraction and modeling of global–local features to accomplish three winding diagnostic tasks: fault type classification, degree quantification, and spatial localization. Experimental data and field cases show that the proposed diagnostic framework can effectively perform all winding fault diagnostic tasks, with diagnostic accuracy and F1 score of up to 98.08% or more, which is a significant advantage compared with other mainstream frameworks. This innovative framework enables highly accurate identification of winding faults, contributes to enhanced intelligent diagnostic capabilities of power transformers, and offers an efficient and automated diagnostic approach to help maintain the reliability of power systems.