FEGNN: Graph Neural Network with Feature Embedding for Automatic Modulation Classification

Modeling time series as graph structures and employing graph neural networks for modulation classification demonstrates significant potential. Unlike existing methods that treat sampled points as graph nodes to construct graph structures, this study proposes an automatic modulation classification algorithm, named Feature Embedding Graph Neural Network (FEGNN). Firstly, FEGNN utilizes sequence permutation for data augmentation, incorporating positional information of modulation signals to generate inputs for graph nodes. Subsequently, the augmented signals are fed into a feature embedding network to create fully connected graphs using node embedding features. Finally, modulation classification is performed using graph neural networks. Simulation experiments under two publicly available benchmark datasets reveal that the proposed method outperforms the existing modulation classification approaches based on CNNs, RNNs, and GNNs.