Research On Cross-subject Mental Fatigue Detection Based On Multi-scale EEG Features

Different from muscle fatigue,mental fatigue is a mental state that will lead to drowsiness,lack of concentration,and difficulty in thinking.Under a continuous and severe state of mental fatigue for a long time,a person will suffer a lot.Electroencephalogram(EEG)is considered an important tool to detect fatigue states because of its good accuracy and high temporal resolution.Currently,most mental fatigue detection models are designed for a specific subject.Due to the long duration of fatigue-induced experiments,EEG data is difficult to obtain,so it is important to construct cross-subject models using the EEG data of existing subjects.The recognition accuracy of cross-subject fatigue detection is usually low due to the nonlinear and non-smooth characteristics of EEG signals and large individual variability,which cannot meet the needs of online fatigue detection systems.To address these issues,this paper thoroughly investigates the EEG feature extraction method based on multi-domain feature fusion and constructs a cross-subject fatigue detection model based on a domain adaptation network in transfer learning method.Finally,a fatigue detection system is designed and implemented under a 2-back training task.In this paper,the research will include:(1)A study of the multi-scale time-frequency feature fusion method for mental fatigue detection.Currently,for the feature extraction methods used in mental fatigue detection,many studies extract features from a single domain,which will lead to feature information insufficient.Therefore,this paper investigates the relevant time-domain and frequency-domain feature extraction methods based on EEG.Meanwhile,the importance of each feature is calculated by the feature ranking selection method,and the features with high contribution are selected to fuse.The effective time-frequency information is retained while achieving feature dimensionality reduction,which reduces the negative impact of redundant features.The experimental results demonstrate that the proposed method improves the classification performance compared with other feature extraction methods.(2)A study of cross-subject fatigue detection method based on multi-source domain adaptation network.To address the problem of large differences in feature distribution between the source and target domain in cross-subject fatigue detection,a domain adaptation network in transfer learning is used to achieve feature alignment and reduce the differences in feature distribution.In addition,to address the problem of low classification performance resulting from the multiple source domains and a single target domain in a domain adaptation network,a weighted loss function is used to improve the performance by assigning different weights to the samples from different domains.It is demonstrated that the improved domain adaptation network obtains better classification results.And compared with other mainstream domain adaptation methods,the improved network has better classification performance.(3)Design and implementation of a real-time fatigue detection system based on a 2-back training task.Based on the E-prime platform,a 2-back training task is implemented,which aims to enhance memory.And a real-time fatigue detection system based on the 2-back task is implemented.During the subjects’ performance of the 2-back task,the system acquires and analyzes their EEG signals in real-time,visualizes and analyzes the EEG features in different states,and finally assesses the subjects’ fatigue level.The fatigue detection system is designed to verify the effectiveness of the proposed time-frequency fused feature extraction method and the cross-subject fatigue detection model.