M4: A Multi-Scale Dual-Domain Feature Fusion Transformer for Automatic Modulation Classification
Many deep learning architectures for Automatic Modulation Classification (AMC) operate directly on time-domain I/Q waveforms and do not explicitly provide a fixed spectral summary to the learned backbone. This paper proposes M4, a lightweight Transformer that appends a fixed discrete Fourier transform (DFT)-derived spectral channel (DualFFT) and integrates three additional modules—ResDepth, SwiGLUConv, and large-kernel ConvNeXt (LGK). The architecture is evaluated on three AMC benchmarks and one controlled synthetic emitter-profile classification proxy, and is documented through a progressive model-development study. Because the M0–M4 stages used stage-specific training settings, this trajectory is descriptive development evidence rather than a controlled causal ablation. The designated M4 checkpoints record 92.34% on RM18-9cls, 63.39% on RadioML2016.10a, 94.40% on Sim-V5, and 54.41% on an archived 50,000-sample RadioML2018.01a subset (RM2018-50k). At 0.83 M parameters and 2.19 ms inference latency on an RTX 4080 SUPER graphics processing unit (GPU), M4 provides a compact operating point. Because the archived baselines are single runs with model-specific optimization settings, accuracy differences are reported as descriptive point estimates rather than inferential rankings.