关键词: 雷达信号, 时频分析, 深度学习, 多尺度空间, 注意力机制

Abstract: Addressing the issue of low recognition rates for complex multi-class radar signal modulation types under low signal-to-noise ratios(SNR), a new method based on time-frequency fusion features and multi-scale dual attention network is proposed. By applying three time-frequency analysis methods, namely, smoothed pseudo Wigner-Ville distribution, Fourier synchrosqueezed transform, and Hilbert-Huang transform based on variational modal decomposition, and combining with denoising preprocessing techniques, the radar signals are transformed into three-channel time-frequency feature maps, which significantly enhances the robustness and expressive power of the features. A multi-scale dual attention network is further designed to realize cross-scale information fusion and noise suppression through the multi-scale channel attention mechanism, while the multi-scale spatial attention is used to adaptively perceive the time-frequency structure of the radar signals and integrate the information through gated fusion and residual connection. The experimental results show that the method achieves an overall recognition rate of 98.99% for 12 types of typical radar signal modulation modes under the condition of a signal-to-noise ratio of -10 dB, showing good robustness.

Key words: radar signal, time-frequency analysis, deep learning, multi-scale space, attention mechanism