Cyclostationarity-Driven Multi-Scale Convolutional Attention Network for Robust Modulation Recognition in Low-SNR Scenarios

Blind modulation recognition, as the core technology in cognitive radio systems, demonstrates critical application value in wireless spectrum management and electronic countermeasures. To address the challenge of blind modulation recognition under low SNR conditions, this paper proposes a hybrid approach that integrates deep features from cyclic spectrum/squared cyclic spectrum slices with handcrafted features. Building upon conventional cyclic spectrum analysis, we introduce squared cyclic spectrum processing to construct a specific feature vector. A multi-scale convolutional attention network (MSCAN) is developed to automatically extract deep features from spectral slices, which are subsequently fused with the handcrafted features for final classification. Experimental results demonstrate that the proposed method achieves an average recognition accuracy exceeding 90% at -7 dB SNR for various modulation schemes, including BPSK, 2FSK, 4FSK, 2ASK, MSK, QPSK and 16QAM, while expanding the types of recognizable signals compared to conventional cyclic spectrum-based algorithms.