Automatic Modulation Classification Based on Multi-Scale Graph Neural Network with LSTM

Automatic modulation recognition (AMR) utilizing deep learning has received considerable attention, demonstrating remarkable potential in extracting complex signal features. However, deep learning-based methods often neglect the spatiotemporal correlation characteristics of signals, which are crucial for signal modulation recognition. Graph neural networks (GNNs) have recently been proposed as a solution to this challenge by converting timr series data into graph datas. In this paper, we introduce a Multi-Scale Graph Neural Network (MSGNN) for automatic modulation recognition. This approach reveals intricate representation within time-domain signals. Specifically, we first leverage Long Short-Term Memory (LSTM) model to adaptively map the time series into visibility graphs. Secondly, a multi-scale graph neural network is designed for modulation classification, which fuses features of different resolutions. We validate the effectiveness of MSGNN on RML2016.10b dataset, demonstrating that our model achieves remarkable recognition accuracy even at low signal-to-noise ratios. The experimental results indicate that MSGNN outperforms the majority of deep learning methods and demonstrates exceptional performance in automatic modulation recognition tasks.