When diagnosing the online operational status of power transformers, the performance and efficiency can be further improved. In light of the extracted vibration signal of the box wall, this paper proposes a detection method which is based on the enhanced gray wolf optimization variational mode decomposition deep belief network (VMD-DBN). Firstly, the energy error is used as the fitness function of the enhanced gray wolf algorithm, and then, all the necessary parameters of the VMD (the number of decomposition levels k and the penalty factor α) are optimized. Additionally, the intrinsic mode functions (IMF) are decomposed and calculated to form a characteristic data set to characterize the operating conditions of the transformers. Finally, the characteristic data set is trained by the DBN, and a fault diagnosis model is generated to identify the operational status of the transformers. The case study results show that the proposed VMD can extract the effective features in the signal better and improve the accuracy of diagnosis. Meanwhile, compared with the other two recognition algorithms, the DBN is better in feature description, learning, and robustness. It can accurately identify four statuses of the transformers, including the normal, the winding radial deformation, the winding axial deformation, and the iron core failure. The state recognition rate of enhanced gray wolf optimization VMD-DBN reaches 97.45%, and the mean error is 0.37, which is the best compared to other methods. Therefore, the proposed method has certain practical value.