Aimed at the problem that the traditional characteristic parameters have difficulty in characterizing the specific characteristics of complex system radar signals, a specific emitter identification algorithm based on deep belief network feature extraction is proposed on account of the deep feature extracting and highdimensional data processing ability of deep belief network. Firstly, a DBN model based on multilayer restricted Boltzmann machine is established. Then, unsupervised extraction of pulse envelope frontier is realized via deep belief network. After that, the model parameters are finetuned with label data in a supervised way to complete the training. Finally, the pulse envelope frontier features of the unknown source signals are input to realize the radar specific emitter identification. Compared with the traditional algorithm, the novel algorithm can adaptively extract from deep pulse features, and can also reduce the process of feature extraction to the dependence on human experiences. The experimental results show that the proposed algorithm provides satisfactory performance of pulse envelope feature extraction and higher recognition accuracy. The validity and application value of the algorithm are verified.