Recognition of LPI Radar Waveforms via RCMNet in Low SNR Scenarios

Abstract

This letter designs a radar convolution-matrix long short-term memory (mLSTM)-based network (RCMNet) for low probability of intercept (LPI) radar waveform recognition. At the modelling stage, the proposed RCMNet architecture operates directly on time-domain I/Q data while incorporating handcrafted interpretable features (including the amplitude and phase) as auxiliary inputs to provide shallow priors for initial signal interpretation. To address the low-SNR challenge, an integrated denoising mechanism is designed in RCMNet, which employs a joint training strategy aiming to optimize both reconstruction loss and cross-entropy loss. Numerical results demonstrate that the devised RCMNet achieves an average recognition accuracy of 88.17% across 12 types of radar waveform at SNR = −15 dB.