To address incomplete transformer winding fault protection mechanisms and low fault diagnosis accuracy

this paper proposes a transformer winding fault diagnosis method based on leakage magnetic field characteristics. Firstly

to avoid interference from noise in leakage magnetic detection affecting the judgment of fault states

an improved Sequential Variational Mode Decomposition (SVMD) method is proposed. This method optimizes its hyperparameters through the Newton Raphson Based Optimizer (NRBO) and combines wavelet thresholding to remove residual noise from the relevant modes after SVMD decomposition. Secondly

the Markov Transition Field (MTF) converts the denoised leakage magnetic signals into two-dimensional images. Fault diagnosis is then performed using a residual neural network (ResNet50) and Multi-Head Attention (MA). Finally

dynamic model experiments are conducted to validate the method

showing that the proposed denoising method effectively restores the leakage magnetic field signals. Compared with four traditional schemes

including Support Vector Machines

the proposed neural network scheme better extracts leakage magnetic field characteristic information

achieving a fault diagnosis accuracy of 96.67%. This demonstrates the effectiveness of the proposed method for transformer winding fault diagnosis.