An improved multi-source domain adaptation network for inter-subject mental fatigue detection based on DANN.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) is a valuable tool for real-time, objective mental fatigue detection.
• Individual variability in EEG signals necessitates time-consuming calibration sessions.
• Existing methods struggle with inter-subject differences in mental fatigue detection.

Purpose of the Study:

• To propose a multi-source domain adaptation network for inter-subject mental fatigue detection.
• To address the challenge of individual variability in EEG data.
• To improve classification performance in mental fatigue detection without extensive calibration.

Main Methods:

• Utilized power spectrum density from four sub-bands of EEG signals, extracted via the Welch method.
• Developed a focal loss based domain-adversarial training of neural network (FLDANN) for feature extraction.
• Employed adversarial learning to minimize feature discrepancies between source and target domains.
• Incorporated focal loss to assign differential weights to source and target domain samples.

Main Results:

• The proposed FLDANN method achieved 84.10% ± 8.75% accuracy on the SEED-VIG dataset.
• The method attained 65.42% ± 7.47% accuracy on a self-designed dataset.
• FLDANN outperformed existing state-of-the-art domain adaptation methods in inter-subject mental fatigue detection.
• Classification accuracy of domain-adversarial training of neural network (DANN) stabilized as the number of source domains increased.

Conclusions:

• The proposed FLDANN method effectively mitigates issues arising from individual differences across subjects.
• FLDANN successfully enhances classification performance in multi-source domain transfer learning for mental fatigue detection.
• The study demonstrates a viable solution for accurate and efficient mental fatigue detection using EEG.