Unknown radar signal recognition based on deep metric learning

Abstract

In view of the fact that the current radar signal modulation recognition method based on depth learning needs enough samples to achieve good recognition effect, this paper designs a depth metric learning recognition method based on Cosine Softmax (CS Softmax), which has strong intra class aggregation and inter class separation, and can recognize unknown signals. First, the Wigner Ville distribution (WVD), Pseudo Wigner Ville distribution (PWVD) and Smooth Pseudo Wigner Ville distribution (SPWVD) of known signals are taken as the input, the vectors obtained from the network output are learned into the embedded space to identify radar signals, and several unknown signals are combined with the initial training set in the verification phase. The simulation results show that under the signal to noise ratio of 10<i>dB</i>, the recognition accuracy of the known signal of this method can reach about 94%, can distinguish the unknown signal from the known signal, and can distinguish the unknown signal separately, with an accuracy of about 85%.