Real-complex Hybrid Network with Fourier Analysis Network for Automatic Modulation Recognition

To address the challenges of insufficient feature representation capability and high model complexity in automatic modulation recognition algorithms under complex channel environments, this paper proposes a real-complex hybrid network architecture based on Fourier analysis. The model innovatively constructs parallel branches of complex-valued depthwise separable convolution and real-valued convolution, dynamically generating joint time-frequency representations through a differentiable short-time Fourier transform (STFT) layer, and achieves adaptive fusion of time-frequency-spatial three-dimensional features using multi-head attention mechanisms. Experimental results demonstrate that the proposed architecture achieves 62.9% average recognition accuracy with only 0.218M parameters on the RML2016.10a dataset, representing a 33.9% reduction in parameters compared to existing state-of-the-art models while reducing error rates by 37.5% under low signal-to- noise ratio (-20dB) scenarios.