Aiming at the shortcoming of traditional mechanical fault detection methods based on vibration signals that multiple feature quantities need to be selected, this paper introduces a complete ensemble empirical mode decomposition (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, CEEMDAN) based on adaptive white noise. The method of judging the mechanical fault of distribution transformer with gramian angular field (GAF). This method uses CEEMDAN to reconstruct the signal and applies GAF transformation to obtain a two-dimensional image of the reconstructed signal. After the two-dimensional image is gray-scaled and binarized, the resulting binary matrix is used to train the radial basis function (Radical Basis). Function (RBF) neural network to realize the detection of mechanical faults. A transformer was used to simulate and test the fault, and the results show that the method is accurate and effective. In engineering practice, the classification function of the RBF neural network can be optimized by continuously collecting a large number of transformer operating data, which can realize the accurate identification of different types of faults, which has high reference value.