EEG-Based Driver Fatigue Detection with Frequency–Spatial Attention and Mamba2 Temporal Modeling

Driver fatigue is a significant research topic in the field of traffic safety. How to utilize electroencephalogram (EEG) signals to rapidly and accurately identify a driver's fatigue state is of great importance for reducing traffic accidents. This paper proposes a network model called CFAMamNet, which combines channel-frequency attention and a MobileNetV3-style lightweight convolutional backbone, and employs Mamba2 for time-series modeling. First, the raw EEG data are integrated in the frequency and spatial dimensions, and a channel-frequency attention module is constructed to capture global interactions among different spatial electrode positions and frequency bands. Next, a convolutional network and multi-branch context module are combined as the backbone to extract and fuse multi-scale spatial features. Finally, Mamba2 is adopted to model the temporal dimension, fully exploring both short-term and long-term dependencies in EEG. We conduct experiments and ablation studies on the SEED-VIG dataset. The results demonstrate that, compared to several existing approaches, the proposed model achieves higher accuracy and greater robustness, confirming its effectiveness and scalability for driver fatigue detection.