The internal leakage magnetic field of transformer is an important criterion for determining the early fault of transformer winding. In actual operation, noise can interfere with the detection of the leakage magnetic field, thereby affecting the judgment of the fault status. Therefore, firstly, genetic algorithms are used with sample entropy as the fitness function to optimize the parameters of variational mode decomposition (VMD). Subsequently, the relevant modes obtained from VMD are processed using wavelet thresholding to eliminate residual noise. Next, feature vectors are selected and extracted from the denoised leakage magnetic field signals. These feature vectors are then input into an improved extreme learning machine (ELM) for training and classification, achieving early fault diagnosis of transformer windings. The results of simulation and dynamic experiment show that this method exhibits a good denoising performance, effectively restoring the original leakage magnetic field signal. Ultimately, it enables accurate identification of early faults in transformer windings.