About radar emitter signal identification research, the artificially extracted features have relatively physical characterization, but there are still redundant features and noise features. Through the deep neural network, the deeper expression of the signal can be obtained, but its characteristics are difficult to explain. Combining the physical characteristics of artificial features and the strong learning ability of deep learning, this study proposes to apply a deep feature selection network to radar signal recognition technology. DFS adds a sparse one-to-one layer between the input layer and the first hidden layer to obtain the corresponding weight value of each feature from the classification correlation metric, uses these weight values to enhance the input of sensitive features and weaken the input of redundant features, and improves classification accuracy. Firstly, the complexity features, Cscade Connection features of ridge-frequency, and information entropy features are extracted from the radar signals, and merged into the original feature set. The DFS is used for learning training to achieve the feature selection at the input level. The above approaches were used to identify the 5 different types of radar emitter signals, obtained good classification. The results verify the effectiveness of the approach.