Abstract: Existing radar emitter signal recognition methods based on manually extract features have problems including low timeliness and poor recognition rate. To address these issues, a new recognition method based on a convolutional neural network with multiple learning units is proposed. First, the burr and distortion caused by noise of ambiguity function of emitter signals are corrected through the Gaussian smoothing. Then, the orthogonal slice is extracted as the further feature extraction objects. Finally, a convolutional neural network with multiple learning units is built to learn and extract the deep and ubiquitous features of the orthogonal slice, which are further classified through the softmax classifier. Simulation results show that the overall average recognition rate of six typical radar signals are all above 99.86% when the signal-to-noise ratio is -2 dB. The recognition rate can reach up to 88.50% when the signal-to-noise ratio is -6 dB. The results prove the good performance and feasibility of the proposed method when signal-to-noise ratiois extremely low.

Key words: radar emitter signals, ambiguity function, signal recognition, deep learning, convolutional neural network with multiple learning units