Noise-Resilient Graph-LSTM Transformer for Enhanced Modulation Recognition

The advent of increasingly complex communication networks necessitates the availability of robust Automatic Modulation Recognition techniques that can adapt to diversity and noisy environments. This paper propose a novel architecture called Noise-Resilient Graph Transformer-LSTM (NRGTL), designed to leverage the unique graph structure inherent in radio signals while obviating the shortcomings inherent in prior methods. Unlike traditional AMR models, NRGTL integrates GNNs for spatial feature extraction with Bi-directional LSTM for capturing sequential dependencies, thus combining the best of both worlds into a hybrid framework that captures both spatial correlations and temporal dynamics in modulation signals. To further enhance noise resilience, a multi-scale attention mechanism is introduced, allowing the model to focus on critical signal components even under extreme low Signal-to-Noise Ratio (SNR) conditions. Adaptive edge pruning is also used in the architecture to refine graph connectivity dynamically, thereby improving computational efficiency without losing accuracy. Results from an extensive evaluation performed on the RML2016.10a dataset show how NRGTL obtains superior results compared to those of the current best state-of-the-art models: Hybrid GNN-Transformer and other LSTM, all at multiple SNR values, for F1-score, Bit Error Rate, and accuracy in general. This work shows the future direction of more progress toward AMR via graph-driven approaches that would be very easily integrated into temporal models for realworld applicability.