MoEformer: An Input-Adaptive Multi-Scale Mixture-of-Experts Transformer for Automatic Modulation Recognition

Automatic modulation recognition (AMR) is essential for spectrum sensing and non-cooperative wireless monitoring. Although multi-scale temporal representations provide complementary information from received in-phase and quadrature (I/Q) signals, their fusion faces two challenges: mismatched feature sequence lengths and the inability of fixed fusion rules to adapt scale contributions to individual inputs. To address these challenges, we propose MoEformer, a multi-scale Mixture-of-Experts Transformer with input adaptability for AMR from raw I/Q sequences. Parallel experts extract scale-specific features from temporally resampled inputs. Frequency-domain feature alignment maps the expert features to a common sequence length, while a learned gating network assigns sample-dependent weights before Transformer-based temporal modeling and attention pooling. Experiments on RadioML2016.10a, RadioML2016.10b, and RadioML2018.01A achieve overall classification accuracies of 63.74%, 66.24%, and 64.22%, respectively, outperforming all compared baselines. For SNR from $-20$ to 0 dB, MoEformer exceeds the strongest baseline on each dataset by 1.07, 1.56, and 1.43 percentage points, respectively. Ablation studies support the contributions of adaptive expert weighting and frequency-domain feature alignment, while the additional inference overhead highlights the trade-off between recognition accuracy and computational efficiency.