Attention-Guided Multi-scale Convolutional Neural Network for Driving Fatigue Detection

Driver fatigue poses a significant safety risk in road transport, prompting heightened public awareness. Electroencephalography (EEG), a key physiological signal for fatigue characterization, has garnered substantial interest and served as de facto gold standard. Existing EEG-based fatigue detection methods often utilize limited forehead lobe channels or the entire brain’s EEG signals, potentially leading to information loss or redundancy. This oversight neglects the crucial aspect of optimizing channels. Moreover, many studies prioritize augmenting model performance through additional sub-networks, neglecting refined features, deepening network and creating overfitting. To address these gaps, this paper introduces an attention-guided multi-scale convolutional neural network for driving fatigue detection, comprising a multi-scale convolutional neural network (MSCNN) and a channel attention mechanism (CAM) module. The MSCNN optimizes network depth, capturing richer temporal and frequency features. The CAM module supplements the MSCNN by emphasizing valuable information, exploring channel contributions, and assessing the impact of feature refinement on model performance. The proposed method demonstrates promising results on the public SEED-VIG dataset with an average root mean square error (RMSE) of 0.051 and an average correlation coefficient (COR) of 0.966 across 23 subjects, outperforming alternative approaches in driving fatigue detection. These findings show that our proposed method perform more effectively for driving fatigue detection based on EEG.