A new triple correlation function–sparse autoencoder (TCF–SAE) algorithm based on SAE for m-sequence recognition is proposed. First, the peak characteristic of the TCF of m-sequences is introduced. The peak characteristic is found to be well kept irrespective of periodic or aperiodic m-sequence. Second, a construction method of input sample for network based on the TCF characteristic of m-sequence is proposed. Finally, a feature learning network is constructed by a SAE, and the learned features are classified by softmax regression. A network model with optimal recognition performance is then obtained by simulation experiments with different numbers of hidden layers and hidden units. The results show that the proposed TCF-SAE algorithm for the m-sequence classification is effective and displays a good recognition performance at low signal-to-noise ratio.