A Dual-Branch Spectral–Temporal Attention Fusion Network for EEG-Based Driving Fatigue Detection

Driving fatigue is a major contributing factor to road traffic accidents, particularly under prolonged driving conditions where it significantly impairs attention and reaction capabilities. Electroencephalogram (EEG) signals, due to their high sensitivity to mental states, have been widely adopted for fatigue detection and assessment. However, existing methods still struggle to jointly model discriminative spectral patterns and multi-scale temporal dependencies. To address these limitations, we propose Dual-Branch Spectral-Temporal Attention Fusion Network (STAFNet) for EEG-based driving fatigue detection. The spectral branch uses frequency-band convolution and squeeze-and-excitation attention to extract key rhythms, while the temporal branch employs multi-scale convolution, Bidirectional Gated Recurrent Unit (Bi-GRU), and temporal attention to capture fatigue-related temporal dynamics. Semantic-level feature fusion is then performed to integrate the two branches collaboratively. Extensive Experiments conducted on a self-constructed 64-channel EEG dataset comprising 40 participants demonstrate that STAFNet outperforms mainstream baseline methods in terms of classification performance and achieves strong cross-subject generalization ability. Ablation studies confirm the contribution of each module, and frequency importance analysis highlights the dominant role of alpha and theta bands in fatigue representation. Moreover, cross-dataset evaluation on the public SEED-VIG dataset further confirms its robustness. These results suggest that STAFNet provides a robust and effective signal processing framework for real-time fatigue monitoring in driving scenarios.