To improve the effect of transformer fault diagnosis，a transformers fault diagnosis method based on weighted comprehensive loss optimization deep learning and dissolved gas-in-oil analysis （DGA） is proposed. The presented method takes DGA characteristic variables as input and probability distribution of fault states in Softmax layer as output，and builds a transformer fault diagnosis model based on stack sparse auto-encoder （SSAE） deep learning. In order to solve the problem that the transformer fault diagnosis effect is low under the normal cross entropy loss function，and the unbalanced distribution of training samples affects the fault diagnosis effect，the weighted comprehensive loss function is used to optimize the deep learning model. The application results show that compared with the traditional methods，the presented method can reduce the adverse effect of training sample asymmetry on transformer fault diagnosis and improve the level of transformer fault diagnosis. The accuracy of the method in this paper can be maintained above 90% for each training set.