Lightweight Recognition of LPI Radar Signals Based on SHNet

To address the problems of low accuracy, high algorithm complexity, and large computational load in signal recognition for Linear Frequency Modulation (LFM) radar signals under low Signal-to-Noise Ratio (SNR) conditions, this paper introduces a algorithm that combines the Smooth Pseudo Wiener-Ville Distribution (SPWVD) with a Convolutional Neural Network (CNN). The algorithm starts with a ShuffleNetV2 network model, incorporating the CBAM attention mechanism to make adaptive adjustments to the weights of distinct channels in the feature map. To reduce the computational load, number of parameters, and overfitting risk, the number of stacks in the basic units is appropriately reduced. To acquire global information in the dimensions of both height and width, a Global Grouped Coordinate Attention Module (GGCA) is added after the maximum pooling layer of the network. This results in a new convolutional neural network model called SHNet. Finally, time-domain signals with varying SNR noise are transformed into two-dimensional images through SPWVD time-frequency transformation and fed into the SHNet network model to complete the radar signal recognition process. The results of the experiments indicate that the recognition accuracy of this approach can attain more than 80% at − 12dB low SNR, which improves by 3 percentage points, and meets the requirements of lightweight recognition of LPI radar signals at low SNR.