Abstract:Recently, radar signal recognition methods based on deep learning have achieved great success. However, these approaches still need to be improved due to the limitation of the number of parameters and computation complexity. This paper proposes an intra-pulse modulation recognition approach for low probability of intercept (LPI) radar signal based on time-frequency analysis (TFA) and lightweight convolutional neural network. Through TFA and image preprocessing, the time-frequency images (TFIs) of radar signals are extracted, noise interference is reduced and redundant frequency band is removed. In order to maintain resolution of images and enhance the ability of the network to extract channel and location information, we introduce dilated convolution and coordinate attention (CA) to ShuffleNetV2. And we proposed a joint loss function consisting of center loss and label smoothing loss to alleviate the overfitting problem and make samples cluster better. Simulation results show that the proposed approach can achieve an overall accuracy of recognition of 98.14% when the SNR is -8 dB.

LPI Radar Signals Intra-Pulse Modulation Recognition Based on CA-ShuffleNet