Wavelet convolutional neural network and dual-channel feature extraction for automatic modulation classification

Automatic Modulation Classification (AMC) identifies modulation formats to ensure correct signal demodulation and interference signal recognition for wireless communication. To address the issues of feature degradation and high misjudgment rates in AMC under low Signal-to-Noise Ratio (SNR) conditions, this paper proposes a method combining a wavelet convolutional frontend, heterogeneous dual-channel feature extraction, and Squeeze-and-Excitation (SE) reweighting. First, a multiscale Wavelet Convolutional Neural Network based on a fixed Morlet wavelet kernel is constructed to perform bandpass analysis on the input, suppressing out-of-band noise while preserving discriminative features in the carrier frequency neighborhood. Second, a dual-channel feature extraction structure is designed, incorporating IQ sequences into Convolutional Neural Network (CNN) and AP data into Long Short-Term Memory Network (LSTM) branches separately, with SE module after each branch for sample-adaptive channel reweighting. Finally, features are fused to achieve complementary spatial and temporal information before feeding into the classifier. Experiments on the RadioML2016.10a, RadioML2016.10b, and RML22 datasets demonstrate that the proposed method exhibits strong adaptability across different datasets with moderate complexity. At SNR = -4 dB and 0 dB, the classification accuracy reaches 87.5% and 93.7% (RML22), respectively, while the highest accuracy achieves 96.8% (RML22) at SNR = 18 dB. It outperforms six baseline comparison models, particularly under low signal-to-noise ratio conditions, with the maximum improvement in accuracy reaching up to 19.0%(RadioML2016.10a) at SNR = -6 dB. The proposed approach provides practical frontend design and model integration techniques for low-SNR wireless signal demodulation and interference identification in engineering applications.