LSTM Model with Self-Attention Mechanism for EEG Based Cross-Subject Fatigue Detection

Abstract

The Brain-Computer Interface (BCI) system based on electroencephalography (EEG) is proven to detect human mental fatigue. In recent approaches, however, the procedure of EEG data requires a lot of feature engineering and is challenging to achieve ideal recognition accuracy in cross-subject scenarios. This paper presents a novel deep learning model towards remarkably accurate based on the self-attention-based Long Short-Term Memory (LSTM) model. Our study shows that LSTM can find the relevant features of each frequency band between acquisition channels, rather than independently concatenate high-dimensional EEG data into a feature vector; the self-attention mechanism can select the information that is more critical to the current mission from high-dimensional data. In the experiment, the public dataset we selected was labelled with two fatigue levels, and the sample balance was achieved by randomly deleting most samples. Our result shows that our model achieves a 78.84% accuracy rate and outperforms the other methods in a cross-subject situation for fatigue detection. Specifically, self-attention based LSTM improves the accuracy higher than EEG-Net by 19.84% and subject matching by 4.52%.