MLEA-AMR: A Multi-Domain Lightweight Expert Framework With SNR-Conditioned Attention for Robust Automatic Modulation Recognition

Abstract

Automatic modulation recognition (AMR) is a fundamental task in cognitive communications and intelligent spectrum monitoring. Existing AMR methods often underexploit complementary signal representations or rely on computationally expensive backbones. This paper proposes MLEA-AMR, a lightweight SNR-conditioned expert aggregation framework. MLEA-AMR constructs three deterministic views from the same normalized complex baseband sequence, including time-domain features, short-time Fourier transform (STFT)-spectrum representations, and multi-interval differential polar (MIDP) descriptors. Compact domain-specific experts extract representation-adapted features, while an SNR-conditioned fusion module aggregates expert outputs through SNR-prior routing, cosine attention, and LayerScale stabilization. The SNR input is treated as a noise-regime condition rather than channel-state information. Under a unified class-SNR stratified evaluation protocol, experiments on RadioML2016.10a, RadioML2016.10b, and RadioML2016.04C show that MLEA-AMR achieves a favorable accuracy–efficiency trade-off with only 0.34M parameters. Additional ablation, SNR-uncertainty, and inference-complexity analyses further validate the effectiveness and practical scope of the proposed design.