Low Probability of Intercept Radar Signal Classification using Multi-Channel 1D-CNN

Modern radar systems are designed to emit low probability of intercept (LPI) waveforms to avoid interception and detection by enemies. In this process, automatic radar LPI waveform recognition becomes a helpful tool for electronic counter-measures. In this paper, we propose LPI-Network (LPI-Net) that uses the complex radar returns to the multi-channel 1-dimensional (1D)-CNN. The direct features of the complex radar signals such as real, imaginary, and absolute values are used as inputs to the multi-channel 1D-CNN network. This LPI radar waveform recognition considers thirteen different waveforms: Barker, Frank, P1, P2, P3, P4, linear frequency modulation, rectangular waveform, T1, T2, T3, T4, and Costas. As a preliminary investigation, we verified various combinations of multi-channel CNNs and observed that a three-channel (real, imaginary, and absolute) and two channels (real and imaginary/real and absolute) are the three suitable candidates for accurate recognition activity. The proposed three-channel LPI-Net is verified with 10 -fold testing validation and achieves $93.94 \pm 0.59 \%$ overall accuracy at $0 \mathrm{~dB}$ SNR. The model is deployed on different edge computing devices such as Raspberry and NVIDIA AGX to test the feasibility of real-time deployment. The proposed model achieves $78.7 \%$ accuracy at $-8 \mathrm{~dB}$ SNR with a model size of 1.9MB and an inference time of 0.41 milliseconds in NVIDIA A100 GPU.