Automatic Modulation Recognition via Pruned LSTM-GRU with Multi-Head Attention

Automatic Modulation Recognition (AMR) is a key technology in non-cooperative communication systems, significantly advanced by deep learning. However, existing AMR algorithms struggle with low accuracy under low signal-to-noise ratio (SNR) conditions, high model complexity, and long training times. To overcome these challenges, this paper proposes a weight-pruning-based LSTM-GRU network with the multi-head attention mechanism (PMH-LSTMGRU). The model integrates fourth-order cumulants of IQ signals with wavelet thresholding denoising to improve anti-interference capability and SNR. Key features are selected via the multi-head attention mechanism, while weight pruning reduces computational complexity. Experimental results show the model delivers excellent modulation recognition performance across various SNRs. Weight pruning further enhances convergence speed and training efficiency, highlighting its practicality in complex communication environments. This architecture offers an efficient solution for AMR in non-cooperative communication systems, with significant theoretical and practical value.