Dual-stream cross-attention network with fusing local recursive and global evolutionary features for drone RF signal classification
Wuhao Li, Siming Li, Yiqing Tan, Liufan Chen, Yong Qiang JIA · International Journal of Parallel Emergent and Distributed Systems · 2025
This paper proposes a novel unmanned aerial vehicle (UAV) radio frequency (RF) signal classification method using a dual-stream cross-attention network that fuses local recursive and global evolutionary features. It generates recurrence plots and gramian angular summation field (GASF) plots through phase space reconstruction and polar coordinate mapping, capturing local and global signal characteristics. A dual-stream architecture combining convolutional neural network (CNN) and Swin Transformer extracts and integrates these features, enhanced by a cross-attention dynamic gating mechanism (CADGM). Experiments show 98.6% accuracy on 16 UAV RF signal classes, outperforming ResNet50 and Vision Transformer (ViT) by at least 4.1%.