Hybrid attention-based temporal convolutional networks for driver fatigue detection using physiological signal

Fatigue detection remains a major challenge in safety-critical domains such as transportation and aviation, where timely and reliable monitoring is essential to prevent accidents and performance degradation. This study introduces a novel dual-architecture hybrid framework that leverages deep temporal modelling and attention mechanisms for real-time EEG-based fatigue classification. Specifically, two enhanced Temporal Convolutional Network (TCN) variants, TCN with Self-Attention (TCN-SA) and TCN integrated with Transformer Encoders (TTCN), are developed to capture both short- and long-range dependencies in EEG signals. To enhance decision-making robustness, high-level features extracted by these deep models are further classified using traditional machine learning algorithms, including Support Vector Machines (SVM), k-Nearest Neighbours (KNN), Random Forests (RF), and Multi-Layer Perceptrons (MLP). Experimental results demonstrate that the hybrid TTCN+RF model achieves superior performance, reaching 98.47% accuracy in participant-wise evaluation and 93.25% in aggregated analysis. The proposed models are rigorously validated using 10-fold cross-validation repeated over 10 runs and leave-one-subject-out (LOSO) analysis. Furthermore, comprehensive statistical evaluations, including Friedman and Wilcoxon signed-rank tests, confirm the significant performance advantage of the proposed hybrid models across datasets. These findings underscore the potential of hybrid deep learning frameworks for real-time cognitive monitoring in high-stakes environments. • Proposed hybrid TCN-SA and TTCN frameworks for EEG-based fatigue detection. • Dual-architecture design captures short- and long-range temporal dependencies. • TTCN+RF reached 98.47% accuracy, outperforming EEGNet, EEGConformer, and EEGTransformer. • Model learns discriminative features directly from raw EEG, avoiding manual extraction. • Statistical tests confirmed significant gains over baseline and SOTA models. • Enables real-time EEG fatigue monitoring in safety-critical environments.