Knowledge Embedded Convolutional Transformer Hybrid Network for Automatic Modulation Classification

Deep learning (DL)-based automatic modulation classification (AMC) exhibits substantial potential to improve signal processing accuracy and efficiency in intelligent communication systems. However, its practical performance is significantly undermined by severe signal deterioration under complex scenario. Moreover, the DL-based AMC tends to be less efficient at identifying different orders of the same type. To address these challenges, a novel knowledge embedded convolutional-transformer hybrid network (K-CTHN)-based AMC is proposed by integrating artificial features derived from domain knowledge with adaptive features learned via a hybrid convolutional neural network and transformers architecture. Then, a weight screening-based pruning strategy is employed to simplify the proposed network, facilitating its deployment in resource-constrained AMC system scenarios. Simulation results show that the proposed K-CTHN achieves over 98% classification accuracy for signal-to-noise ratios (SNRs) of 0 dB and above, and it outperforms both the data-driven multichannel convolutional long short-term deep neural network and the hybrid data- and model-driven convolution and frequency global filter neural network. More importantly, the classification performance of the same modulation type can be improved even at lower SNR, especially the classification accuracy of the common quadrature amplitude modulation (QAM) family can be kept above 90%. Furthermore, the lightweight version not only reduces the parameters by 63% and computational cost by 60%, but also achieves comparable accuracy to the K-CTHN model.