Research On Attention Analysis Method Based On Brain-computer Interface

Attention level directly determines the efficiency of study and work.Brain-computer interface(BCI)is an important way of human-computer interaction,which provides a new direction for the study of attention.In the attention analysis,we use the Power Spectral Density(PSD)of the five bands ??????? and ? as the feature vector for attention recognition,and the classification accuracy is relatively stable but not high.How to select the most suitable feature for each subject from the PSD features is the key to improve the accuracy of attention recognition.A lot of time must be spent on offline experiments before online experiments in order to obtain EEG data for training models.It is the focus of this paper to ensure the classification performance of online experiments without conducting offline experiments.For the problem that using all PSD features for classification results in low accuracy,This paper designed genetic algorithm(GA)for feature selection.Using intra-class and inter-class distance and Bhattacharyya distance as the fitness function of the GA,the experimental results proved that the Bhattacharya distance has the more correlation with the classification performance.In order to solve the problem of low efficiency and possible elimination of good individuals in the process of algorithm operation,this paper adopted a sub-protection mechanism for individuals with greater adaptability.Individuals in the top 20% of fitness were directly inherited to the next generation,and individuals in the top 20% to 50% of fitness determined whether to inherit or not according to conditional probability strategies.In the selection of classifiers,support vector machine(SVM)and back-propagation(BP)neural network were used for classification respectively.SVM classifier showed higher classification accuracy and classification efficiency.Finally,the experimental results of 10 subjects showed that the genetic algorithm with sub-protection mechanism can significantly improve the classification performance of attention.For the problem that before online experiments must spend a lot of time on offline experiments,this paper introduced a transfer learning algorithm based on adaptive probability distribution to reduce the difference between the source domain and the target domain,and then cancel offline experiments.In order to improve the accuracy of transfer learning,this paper proposed three improvements: First,we removed features with larger Wasserstein Distance to initially reduce the difference in probability distribution between the source domain and the target domain.Second,we used the voting principle to predict the target domain in order to get the pseudo-label of the target domain.Third,in order to obtain a more accurate MMD distance matrix,we introduced weight coefficients based on Wasserstein Distance to dynamically adjust the proportions of the marginal distribution and the conditional distribution.The experimental results proved that the improved probability distribution adaptive algorithm proposed in this paper is effective.Finally,this paper designed an online experiment paradigm based on the existing QT development framework.Five subjects were invited to conduct online experiments.The final experimental results showed that the attention analysis algorithm proposed in this paper is also suitable for online attention detection.