To improve the transformer fault diagnosis effect and weaken the adverse effect of unbalanced training samples, a transformer fault diagnosis method based on variational auto-encoders (VAE) preprocessing deep learning and DGA for unbalanced samples is proposed. The DGA characteristics of each sample and the probability distribution of each fault state are used as the input and output of the diagnosis model respectively. Firstly, the minority training samples are preprocessed by VAE, and the training samples are generated automatically based on the learning to determine the distribution characteristics of the minority training samples, so as to improve the balance of the training samples. Secondly, the transformer fault diagnosis model is constructed based on the stack sparse auto-encoder (SSAE) deep learning network with three-hidden layers, and the parameters of the diagnosis model are updated and optimized with the balanced training samples preprocessed by VAE. The effectiveness of the proposed method is verified based on examples. The experimental results show that the proposed method can improve the adverse effects of unbalanced training samples. Under each training set, the accuracies of diagnosis results using the proposed method maintain above 91%, and the false negative rate are relatively low.