It is of great practical significance to study the fault diagnosis of transformers for safe and stable operation in power systems. The traditional transformer fault diagnosis method with dissolved gas characteristics in oil as input has a large limitation in dealing with sample imbalance data. To address this problem,a transformer fault diagnosis method based on Focal loss-stacked sparse denosiing auto-encoder (SSDAE) is proposed. The method is used to determine hyperparameters by category weights and adds Gaussian white noise as the original input,which facilitates the self-encoder to fully extract effective features,and thus obtaining an effective deep feature extraction model. The Focal loss function is used to optimize the model and a Softmax classifier is used to output the diagnosis results. The results of the case study show that compared with traditional transformer fault diagnosis methods such as three-ratio method,back propagation neural network (BPNN) and support vector machine (SVM) method,the method in this paper further improves the diagnosis accuracy.