Improved Radar Signal Recognition by Combining ResNet with Transformer Learning

Modulation recognition of radar signals in time-frequency images (TFIs) under low signal-to-noise ratio (SNR) conditions has important applications such as electronic reconnaissance in electromagnetic environments. Existing work in applying deep learning for radar signal modulation recognition focuses on deploying convolutional neural networks (CNN) to extract signal's local features, but ignores global features and signal position features in the image, resulting in unsatisfactory recognition results under low SNR. This paper proposes an effective neural network based radar signal recognition model that is able to extract both local and global features of time-frequency radar signals more effectively by combining CNN-based ResNet and transformer learning to improve the recognition accuracy for radar signals under low SNR. Specifically, we apply Choi-Williams distribution (CWD) to transform the signal time domain waveform into a two-dimensional time-frequency image and display the intra-pulse characteristics of the signal in both time domain and frequency domain. We perform preprocessing including grayscale and binarization on the time-frequency images to better highlight the signal characteristics. Then, as the central part of this work, we integrate ResNet feature extraction with multi-head self-attention (MHSA) enabled transformer learning to build our ResNet-MHSA model, namely RM-Net, and use the SCR-loss evaluation module to train the model. Experimental results show that when the SNR is -10dB, the radar signal recognition rate of 12 different modulation types can reach 95.6% with our method, which is an increase of more than 6% than existing work.