Aiming at the problem of strong subjectivity and feature redundancy of radar signal recognition in the complex electromagnetic environment, a recognition method based on deep convolutional neural network is proposed. By extracting the bispectrum information of the radar signal as the input of the network model, the network model is used to automatically learn the deep features, identify the different modulation pattern signals, and compare the recognition results of the deep network models with different structures. The results of simulation experiment show that compared with the traditional radar signal recognition method, the proposed method has improved recognition rate and noise immunity.