Research Of Mental Workload Recognition Based On Multi-dimensional EEG Features

Nowadays,more and more physical labor has been replaced by mental labor,and the human-machine operation has taken up most of the content of work.The work efficiency of human-machine operation will be affected by the mental workload state of human body.In order to improve the efficiency of human-machine operation,to achieve rational allocation of human resources and ensure the safety of staff,it is of great significance to study the mental workload state of human-machine operators.Many studies have shown that electroencephalogram(EEG)has a high correlation with the state of mental workload.EEG acquisition has the advantages of high time resolution,non-invasiveness and accuracy,so the mental workload recognition based on EEG has become a hot research direction.Aiming at visual and operational tasks,this thesis studies mental workload recognition methods suitable for real scenarios by collecting EEG of 8 subjects under different mental workload states.Since the collection of EEG is carried out by multi-channel EEG caps and distributed in various frequency bands,the feature dimensionality obtained through feature extraction is too high,resulting in the high complexity of the subsequent recognition model.To solve this problem,the method of feature dimensionality reduction is usually used to reduce the dimensionality of high-dimensional feature vectors.In this thesis,the existing feature dimensionality reduction methods widely used in EEG features are compared to finding a more suitable feature dimensionality reduction method for mental workload recognition in real scenes.In addition,in order to facilitate the unified research on mental workload recognition of multiple sets of EEG data from different subjects,an adaptive dimensional optimization method is proposed in this thesis.The high-dimensional features of the experimental data set are reduced to each dimensionality within the optimized dimensionality range,and the classification accuracy of each dimensionality is plotted as a dimensionality-classification accuracy curve.The optimal dimensionality reduction dimension of the experimental dataset is determined by identifying the "elbow" of the curve.This method can accurately identify the optimal dimensionality reduction common to multiple data sets under the same experiment and effectively improve the recognition efficiency.Furthermore,the practical application of mental workload recognition methods in real-life scenarios often involves a limited number of historical samples for the subjects under investigation,leading to suboptimal classification performance of the trained classifier.To address this issue,this study presents a cross-subject mental workload identification approach based on mult-idimensional EEG features.From two perspectives of instance transfer and feature transfer,a instance transfer method based on instance selection and a feature transfer method based on kernel principal component analysis is proposed.The instance transfer method based on instance selection uses the EEG of other subjects as the training set,and selects the feature instances in the training set by referring to a small amount of historical EEG of the target subjects.By screening out the instances that are closer to the instances of the target subjects in the training set,the number of instances can be reduced and the differences between the training set and the test set can be reduced.The feature transfer method based on kernel principal component analysis realizes the reduction of feature dimensionality by introducing kernel function to nonlinear dimensionality reduction of training set features and test set features,This technique aims to identify a shared low-dimensional feature space between the training set and test set,thereby further reducing the domain discrepancy between the two.The application of instance selection and kernel principal component analysis to the recognition of cross-subject mental workload can not only improve the recognition efficiency of cross-subject mental workload,but also improve the recognition accuracy of cross-subject mental workload.This method can realize rapid and accurate recognition of mental workload state.