Aiming at the self-learning of fault feature and fault diagnosis in power communication network, a fault diagnosis method based on Stack Relational Sparse Autoencoder (SRAE) deep neural network is proposed. Considering that Management Information Base (MIB) data in power communication network is encoded into binary sequence as training data, the self-learning and feature extraction ability of shallow single autoencoder is insufficient. Therefore, the sparsity restriction and input data relational restriction are integrated into autoencoder to form a Relational Sparse Autoencoder (RAE); Then, it is stacked layer by layer, and the classifiers are added after the last hidden layer to form a SRAE; finally, a simulation experiment is carried out with the connection fault between routers. The experimental results show that the average accuracy of the proposed method is 99.625%, which has high and stable diagnostic accuracy. This work is supported by Key Project of Shanxi Province Key Research and Development Plan (No. 201703D111027) and Science and Technology Project of State Grid Shanxi Province Electric Power Company.