Open-Set Recognition of LPI Radar Signals Based on Deep Class Probability Output Network
Xiao Han, Shiwen Chen · 2022
Radar signal waveform recognition is one of the key parts of Radar Emitter Identification (REI). With the developing of low probability of intercept (LPI) radar signals, there have been many researches on LPI signal waveform recognition. Method based on time-frequency analysis (TFA) and Deep Learning (DL) network that can automatically extract features from images is becoming popular. However, the most existing DL-based methods of signal waveform recognition are time-consuming in preprocessing process and more importantly, is invalid in open-set scenario. To solve these problems, this paper proposes an open-set radar signal recognition method named Deep Class Probability Output Network (DCPON) which combines the great feature extraction ability of DL and the interpretability of CPON fitting probability distribution. It can recognize new unknown signal waveforms while classifying known ones and is more reliable than previous DL-based methods. In addition, the one-dimensional residual network (1D ResNet) is used as the backbone network, and the cross-entropy loss and center loss are used for training. We compare the different preprocessing methods and set the real part and imaginary part of Discrete Fourier Transform (RIODFT) of signals as the input of two channels respectively. Several simulated experiments demonstrate that the proposed method has good performance in the open electronic environment.