An efficient method for cross-subject EEG-based mental fatigue recognition

摘要

Mental fatigue has an unignorable impact on people's daily life and work. Mental fatigue recognition methods based on EEG are commonly thought as objective standard. However, the procedure of initiating fatigue needs a long time and EEG signals vary greatly, so mental fatigue recognition based on EEG is challenging when subject-specific data are limited and imbalanced. In this paper, we explored the performance of cross-subject fatigue recognition on a general multitask learning framework with two data sampling methods for imbalanced classification. One was to synthesize some minority samples (MTLSMS) until the two classes were balanced and another was to under-sample the majority samples (MTLUMS). EEG data of 11 subjects from a public EEG fatigue dataset were selected to validate our fatigue recognition methods. The MTLSMS method improved the cross-subject mental fatigue recognition accuracies to 81.07%, and much better than other transfer learning domain adaptation methods Transfer Component Analysis (TCA) and Maximum Independence Domain Adaptation (MIDA). The experiment results showed that the MTLSMS method can effectively recognize the cross-subject mental fatigue.