Abstract: Aimed at the problem of traditional modulation recognition method which is influenced by signal feature parameters designed by artificial experience , a novel modulation recognition algorithm based on Sparse Stacked Auto-Encoder. First, according to the network input data form requirements, In order to take advantage of the signal modulation information contained in the amplitude and phase, A signal preprocessing method is proposed to preprocess the complex signal into a network-acceptable real form。In the training stage, the initialization parameters of each sparse auto-encoder network are obtained by layer-by-layer training, and then the supervised algorithm is used to train the classification layer. Finally, the supervised algorithm is used to overall optimization. Using softmax regression classifier as the classification layer to complete the Modulation recognition. The simulation results of seven kinds of digital modulation pattern recognition show the effectiveness of the proposed algorithm. Compared with other algorithms, algorithm has high recognition rate at low SNR, and the recognition performance is not affected by human factors.