MAFFNet: a multi-modal adaptive feature fusion net for signal modulation recognition

MAFFNet: a multi-modal adaptive feature fusion net for signal modulation recognition

연구 분야: Infrastructure

학회: The Journal of Supercomputing

Automatic modulation recognition (AMR) is a fundamental process between signal detection and demodulation. Despite recent advances in deep learning-based AMR, existing methods often fail to maintain robustness in severe noise scenarios. To address this, we propose MAFFNet, a noise-robust multi-modal architecture that synergistically processes raw in-phase/quadrature (IQ) signals and derived amplitude/phase (AP) information through a dual-branch vision transformer-LSTM framework. Specifically, the modified vision transformer (ViT) branch employs localized attention mechanisms to reduce computational complexity, while the LSTM branch incorporates phase-difference attention to model temporal dependencies in AP information. Additionally, a learnable feature fusion module with element-wise weights dynamically combines multi-domain features, complemented by an orthogonal constraint loss that reduces inter-branch redundancy. Extensive experiments on the RML2018.01 benchmark show that the proposed MAFFNet achieves 82.53 accuracy between 0–10 SNRs, outperforming other methods by 5.46–11.25%.