Research On The EEG Response And Classification Of Mental Workload And Adaptive Brain-computer Interface

Mental workload(MWL)is the occupancy rate of mental rescource under working condition.MWL has a strong impact on human performance.Performance would be diminished and even human error and terrible accident would happen when MWL is too low or too high.Electroencephalogram(EEG)is one of the most important and most potential psychophysiological measure to investigate the neural mechanism of MWL and to recognize the level of MWL.EEG-based method can recognize MWL in portable and real-time way,and thus can be used for real-time monitoring MWL and building adaptive human-machine interaction system which can adjust to MWL in complicated human-machine systems for optimizing the relationship between human and machine.However,there're plenty of challenges still ahead in the current EEG-based MWL studies: 1)EEG-based research on the neural mechanism of MWL,especially the comparative study of neural response caused by MWL across different tasks and the comparative study of sensitivity of different EEG components to MWL variations.2)The chance level accuracy of cross-task and cross-subject MWL classification in current studies,especially in complicated practical situations.3)The effect of mental workload on the performance of event-related potential(ERP)-based brain-computer interface(BCI)and the countermeasures to cope with it.To address these challenges,the following studies were carried out in this thesis:First,two experiment,a single-task paradigm and a dual-task paradigm,were designed based on cognitive rescources theory to investigate the response of different EEG components to the variation of MWL by analyzing the power spectral density(PSD),the task-relevent ERP(trERP),the task-irrelevent ERP(tirERP),the scource localization of ERP and spontaneous EEG(sEEG),and brain network under very different tasks.Some common patterns of sEEG response to the variation of MWL were found.In particular,the very common response patterns of tirERP to the variation of MWL under very different tasks were firstly found.Second,based on the common responses of sEEG to MWL of different tasks,combining with support vector regression,a cross-task feature selection method was proposed to cope with the challenge of MWL recognition by using PSD features of sEEG.With these approaches,the cross-task challenge was firstly addressed and we realized cross-task MWL recognition between different N-back tasks and between Nback tasks and a complicated simulation task.Third,based on the common responses of tirERP to MWL of different tasks,we further demonstrated that the separability of tirERP features between low and high MWL has much stronger consistency than that of PSD features of sEEG in both crosstask and cross-subject problems.On the foundation of these findings,we firstly realized cross-task and cross-subject MWL classification using tirERP features,and we found that the accuracies of cross-subject classification using tirERP features were batter than that using PSD features of sEEG.Meanwhile,the number of subjects in training data is found to be much more important than the number of training samples.Fourth,on the foundation of the findings that MWL can significantly affect tirERP,the effect of MWL on ERP-BCI was further investigated.We found that the performance of ERP-BCI created under a pure condition would be significantly diminished if it is used under higher MWL conditions.However,training ERP-BCI under appropriate MWL could significantly weaken the negative effect of MWL of testing condition.Based on these findings,a framework of adaptive ERP-BCI was firstly proposed by combining passive BCI and ERP-BCI and its effectiveness in mitigating the impact of MWL on the performance of ERP-BCI was confirmed in online simulation analysis.It means that adaptive BCI is a very promising aprouch to cope with the challenges posed by the sensitivity of BCI to variation of MWL.In conclusion,this dissertation focused on the challenges in current studies and thoroughly analysed the responses of different EEG components to MWL across tasks and subjects,proposed a feature selection method and modeling aprouch for PSD-based cross-task MWL classification,put forward a new approach for cross-task and crosssubject MWL classification based on tirERP,proposed the concept of adaptive BCI and validated its ability in dealing with the challenge confronting the conventional BCIs.