Fatigue detection from single-channel EEG using a hybrid CNN with time-domain and nonlinear feature fusion

Sustained vigilance is a critical requirement in many occupations, but prolonged focus frequently leads to fatigue, which impairs alertness and can compromise operational safety. Therefore, the accurate and unobtrusive detection of operator fatigue is a significant challenge. Although multi-channel Electroencephalography (EEG) is a powerful tool for fatigue analysis, its complex setup hinders practical application in many operational environments. To solve this limitation, this paper proposes a method for fatigue detection utilizing only a single-channel EEG signal. Specifically, we first effectively induced the fatigue state of subjects by employing a psychomotor vigilance task (PVT) experiment and developed a compact EEG acquisition device to collect single-channel EEG signals during the experiment. Subsequently, we designed a novel hybrid two-branch convolutional neural network. One branch applied a Limited Penetrable Horizontal Visibility Graph (LPHVG) to extract the nonlinear dynamic features of the single-channel EEG signals. The complementary branch directly extracts time-domain features from the raw signal. By fusing nonlinear time series analysis with time-domain feature extraction, our method achieved excellent performance on the PVT dataset and maintained robustness even when trained with limited data, demonstrating its effectiveness and practicality.