FFTFormer: A Frequency-Aware Transformer With Multi-Task Learning for Robust Modulation Recognition

Automatic Modulation Recognition (AMR) is a key enabler of intelligent signal processing in modern wireless communication systems. Although deep learning (DL)-based approaches have achieved notable progress, their performance often degrades at low signal-to-noise ratios (SNRs). Furthermore, most existing studies assume fixed roll-off factors during pulse shaping, overlooking the inherent variability of real-world signals, and thus overestimating model robustness. To address these limitations, this paper proposes FFTFormer, a frequency-aware Transformer that integrates a Fast Fourier Transform (FFT)-driven attention mechanism with a frequency-adaptive feedforward network to enhance the learning of modulation features in the frequency domain. By transforming token representations into spectral components, FFTFormer effectively captures modulation-specific spectral dependencies and improves recognition performance in low SNR environments. Moreover, considering the significant influence of the root-raised cosine (RRC) filter's roll-off factor on signal features, an auxiliary roll-off factor prediction task is introduced to guide the adaptive learning of spectral variations across diverse pulse-shaping conditions, thereby enhancing the model's robustness and generalization. Experimental results on the RML16 A and RML22 datasets demonstrate that FFTFormer significantly outperforms state-of-the-art (SOTA) DL-AMR models. Under low SNR conditions, FFTFormer surpasses the best-performing SOTA baselines by 2.36% and 2.83%, respectively. Furthermore, we have verified the effectiveness of the proposed FFTFormer using actual measurement data, emphasizing its practical applicability.