Automatic Modulation Recognition Based on Normalization-Free Transformer

Existing Transformer-based automatic modulation recognition (AMR) methods suffer from computational inefficiency, statistical dependency, and limited feature discriminability. These issues originate from conventional layer normalization's inability to maintain stable feature distributions under non-stationary channel conditions. To address this problem, this paper proposes a normalization-free Transformer architecture incorporating a Dynamic Tanh (DyT) mechanism that adjusts activation responses dynamically through learnable scaling factors. Experiment results on the RML2016.10a and RML2022 benchmark dataset show that the improved models can significantly enhance performance. The DyT-enhanced TLDNN model achieves a peak accuracy of $\mathbf{9 8. 9 \%}$ on the RML2022 dataset with only $\mathbf{0. 3 3 0 M}$ parameters. The average accuracy is boosted by $\mathbf{2. 9 \%}$. Moreover, under a low SNR of $\mathbf{- 4 d B}$, the accuracy improves by 4.4 % relatively. Computational efficiency is also enhanced with the reduced FLOPs by $\mathbf{1. 2 - 1. 3 \%}$.