To address the problem of small samples and unbalanced data of dissolved gas fault data in transformer oil. This paper proposes an enhancement method of dissolved gas data in transformer oil based on the Riemannian mani-fold variational autoencoder. Firstly, the one-dimensional fault samples are transformed into two-dimensional feature images by an improved Gramian Angular Field to construct the original image dataset. Based on the network framework of the variational autoencoder, the latent space is modeled as a Riemannian manifold using the Riemannian metric to capture complex data distributions better; data enhancement of all types of fault samples is realized through the Riemannian Hamiltonian dynamics using the Riemannian Hamiltonian Monte Carlo sampler. The experimental results on the dissolved gas fault dataset in transformer oil illustrate that compared with the traditional method, the approach proposed in this paper can generate more meaningful fault samples without distorting the actual data distribution under the condition of small samples and unbalanced data so that the fault identification accuracy of the deep learning model can be effectively improved.