Multi-scale Graph Neural Network for Low-SNR Wireless Signal Classification

Accurate waveform identification, especially at low SNRs, remains an open challenge; both classical hypothesis-test detectors and modern deep neural networks degrade sharply as SNR falls. We address this with a graph-neural approach that converts each noisy in-phase/quadrature (IQ) segment into a multi-scale temporal graph: every sample becomes a node linked to neighbors at multiple lags (±1,±2,±4,±8) and to feature-space k-nearest neighbors. Each node carries a six-channel feature vector $\left[ {I,Q,|x|,\phi ,\delta \phi ,|x{|^2}} \right]$ designed to surface subtle patterns masked by noise. A hybrid CNN-GNN then performs classification: a 1D convolutional front-end learns local temporal embeddings, followed by a deep residual GNN with interleaved GCN and GATv2 layers to propagate information over the graph. We use Jumping Knowledge (JK) aggregation to preserve multi-hop and shallow cues, and an attention-based readout to emphasize the most informative nodes. On a synthetic four-class dataset (LTE, 5G NR, Radar, noise) at −10dB average SNR, the proposed model attains 91.3% test accuracy, exceeding strong baselines. This demonstrates substantial performance gains over conventional CNNs and deep learning classifiers in the low-SNR regime.