Abstract
EEG-based mental workload (MWL) recognition is challenged by inter-individual variability and limited model interpretability. This study investigated MWL recognition in long-term multi-tasking scenarios using resting-state EEG normalisation and regional feature analysis. Participants performed multi-tasking operations at three difficulty levels, while resting-state and task-state EEG, behavioural performance, and NASA-TLX ratings were collected. The best whole-brain model achieved 93.78% ± 0.45% accuracy under standard normalisation, while the parietal configuration retained substantial recognition capability with fewer electrodes. Resting-state normalisation changed EEG feature-importance patterns, but its effects on classification were classifier-dependent rather than uniformly beneficial. SHAP analysis identified prominent frontal and parietal contributions. The findings support interpretable and lightweight EEG-based MWL monitoring in complex operational environments.