An Attention Neural Network for Radar Intra-pulse Modulation Pattern Recognition

Abstract:
The most existing Low Probability of Intercept (LPI) radar signal recognition algorithms generally have poor recognition performance due to the extraction of radar intra-pulse modulation features at low signal-to-noise ratio (SNR) is difficult. To solve that issue, this paper proposes a novel attention neural network based on the time-frequency images (TFIs) for intra-pulse modulation pattern recognition of radar pulse signals. First, in order to obtain the TFIs from the radar received signals with low cross-terms and high time-frequency resolution, smoothed pseudo Wigner-Ville distribution (SPWVD) is leveraged for generating the TFIs from the received radar signals. Second, Block-matching and 3D filtering (BM3D) is utilized to preprocess the TFIs to eliminate the impact of the background noise for preserving detailed features of the signal, and then we adopt bilinear interpolation algorithm in terms of obtaining the TFIs with the same size for further network training. Finally, the pre-processed TFIs are fed into the deep neural network embedded with SENet attention module, which is called SE-ResNet, to do the recognition. SENet can help the model extract valid feature of TFIs and achieve effective recognition on different radar signal intra-pulse modulation patterns from LPI radar. The simulation results show that the proposed method can achieve an overall recognition rate of 93.2% for eight LPI radar signals when SNR in a low SNR environment, e.g., -8 dB.