Transformer-Based Phase Space Graph Convolutional Network for Modulation Recognition

Recent advances in deep learning-based signal re-representation methods have achieved significant breakthroughs in tasks like signal modulation recognition. In this paper, we propose a Transformer-based phase space graph convolutional network (PSGformer) to enhance performance with a lightweight and high-precision design. PSGformer introduces phase space recurrence networks to map time series to graphs, enabling the extraction of rich signal features using graph data mining techniques. To improve the efficiency of graph mapping, an improved Transformer’s linear attention mechanism is used to parallelize graph construction and convolution, reducing dependency on adjacency matrix thresholds. Additionally, PSGformer integrates a deep shrinkage convolutional network as a side channel to address the sensitivity to Gaussian white noise and linear attention mechanism limitations. Data preprocessing is handled by a frequency selection module. Experiments on multiple datasets (RML2016.10a, RML2016.10b, Sig2019-12, and RML2018.01a) demonstrate that PSGformer outperforms most existing re-representation models.