A Novel Modulation Classification Method Based on Feature Fusion with BiLSTM and Transformer

Automatic modulation recognition (AMR) plays a crucial role in modern wireless communication systems, enabling efficient signal identification and management in complex environments. However, existing approaches face two significant challenges: effectively capturing both spatial and temporal features from raw signals, and addressing the limitations of models that focus primarily on local features while neglecting global feature representations. To tackle these challenges, this paper introduces a novel modulation classification approach that integrates feature fusion with a bidirectional long short-term memory (BiLSTM) and Transformer architecture. By combining spatial and temporal features from in-phase/quadrature (I/Q) and amplitude/phase ($\mathrm{A} / \mathrm{P}$) data of raw signals, the method enhances feature extraction and representation. The feature fusion network (FFNet) captures multi-scale spatial and temporal characteristics, while the Transformer module improves global feature learning. Experimental results on the RadioML2016.10a dataset demonstrate that the proposed method achieves a recognition accuracy of 91.91%, with a peak accuracy of 93.47%. The model also exhibits strong recognition performance under low signal-to-noise ratio (SNR) conditions, maintaining excellent generalization and robustness. These results contribute to advancing automatic modulation recognition in challenging wireless environments.