MSTFFNet: Multi-Scale Time-Frequency Fusion with Self-Estimated SNR Conditioning for Robust Automatic Modulation Recognition

Automatic modulation recognition (AMR) identifies the modulation scheme of received radio frequency (RF) signals under unknown channel conditions and underpins spectrum monitoring and signal demodulation in wireless systems. Under low signal-to-noise ratio (SNR), multipath fading, and limited observation length, however, the discriminative features of modulated signals are severely attenuated, degrading recognition robustness. We propose MSTFFNet, a multi-scale time-frequency fusion network that addresses these challenges with two designs. First, it fuses the raw in-phase/quadrature (I/Q) signal with its short-time Fourier transform (STFT) time-frequency map at the token level through dual-stream heterogeneous encoding, capturing complementary temporal and spectral features. Second, rather than relying on external SNR ground truth, the network self-estimates an SNR-bin probability from the I/Q features and generates a channel-quality embedding that conditions the classifier, requiring no SNR label at inference. On the RadioML2016.10a and 10b benchmark datasets, MSTFFNet achieves overall accuracies of 67.33% and 70.87%, outperforming state-of-the-art methods by 3.53% and 5.33%, with improvements of 5.91% and 9.52% in the low-SNR regime. These results demonstrate improved recognition performance across the SNR conditions represented in the two synthetic RadioML2016 benchmarks, particularly at low SNR.