Few-Shot Radar Modulation Recognition Based on Feature-Level Multi-Time-Frequency Fusion Algorithm

Radar modulation signal recognition plays a critical role in electronic warfare and communication systems, yet the performance of conventional deep learning models significantly degrades in few-shot scenarios due to insufficient training data. To address this limitation, this work presents DB-CBAM-ResNet, a dual-branch attention network that integrates STFT and CWT to extract complementary time-frequency characteristics through a novel feature-level fusion method. The attention mechanism CBAM is employed to adaptively fuse these features through weighted integration, thereby enhancing feature discriminability while reducing dependency on large-scale datasets. Experimental validation on six typical modulation signals (BPSK, QPSK, CW, BFSK, LFM, NLFM) generated by a radar emitter signal simulation platform demonstrates that the proposed method achieves 100 % recognition accuracy in signal environments with SNR <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\geqslant 0 ~\text{dB}$</tex>, demonstrating a 3.0 % improvement over the STFT-only baseline and a 7.3 % gain compared to the CWT-only approach. Notably, it maintains the accuracy of 97 % at SNR <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\geqslant 4$</tex> dB under the extreme training data scarcity of 4 samples per SNR level.