Research On Online Learning Algorithm Based On Brain-computerinterface Of Steady-State Visual Evoked Potential

Brain computer interface(BCI)is a kind of human-computer interaction technology that allows the brain to communicate directly with external computers or devices without relying on peripheral nerves and muscle tissue.Among them,based on the steady-state visual evoked potential(SS VEP)of brain-computer interface system for its non-invasive,signal is relatively stable,high information transfer rate(ITR),easy to train,low equipment price characteristics have been widely studied and applied.As the application becomes more and more widely,the performance of SSVEP-BCI gradually fails to keep up with the development of demand,which becomes the bottleneck restricting the popularization of braincomputer interface.On the one hand,the SSVEP-BCI decoding algorithm cannot meet the application requirements of high performance and low training cost at the same time.On the other hand,in the case of small training samples,the system detection performance is poor,resulting in a long preparation process before each use.To solve these problems,this study took SSVEP detection algorithm as a breakthrough point,supplemented and improved the theoretical model of brain-computer interface signal,and tried to enhance the application potential of SSVEPBCI by reducing training costs and improving data utilization.In order to further reduce the training cost of SSVEP-BCI and improve the system information transmission efficiency,this paper proposes an online learning detection algorithm based on template optimization.Without prior training,the algorithm allows the system to learn the user’s EEG signal in the process of application,which gradually improves the recognition performance from the level of untrained detection algorithm to the level of trained detection algorithm,so as to effectively balance the contradiction between system performance and training cost,and enhance the practical value of the system.Under the condition of no initial training data,the algorithm can recognize the user’s EEG response based on the non-training algorithm.In the process of application,the algorithm can continuously accumulate user unlabeled EEG data,and dynamically update user’s steady-state visual evoked potential response template through online learning,and finally achieve the detection effect with training algorithm.In order to verify the effectiveness of the online learning algorithm proposed in this paper,this study designed a character spelling system based on the joint frequency-phase encoding stimulus paradigm and set up relevant experiments.The results of offline and online experiments on 16 subjects based on 40 target BCI spellers show that the proposed algorithm achieves an average ITR of 145.86±31.10bit/min without initial training data,which is better than CCA,FBCCA,STE-FW and other untrained algorithms.After a period of application,the algorithm achieved the optimal average ITR of 231.55±52.11bit/min when a certain amount of unlabeled training data was accumulated,which was better than the TRCA,eCCA,TRCA-DS algorithm under the same training data.The online experimental results also show that with the expansion of application practice,the information transmission rate of the algorithm can be gradually improved from the initial average 127.19±34.43bit/min to 232.67±66.24bit/min,and the average performance can be improved to 82.93%.In addition,the free spelling experiment also shows that the algorithm has good universality.In order to improve the detection performance of SSVEP-BCI system in the case of small samples,this paper further proposes an optimization of SSVEP detection algorithm based on the idea of metric learning.By learning the class neighborhood,this method establishes a new metric function,which enhances the classification and recognition effect of the original algorithm on SSVEP signals.The off-line experiment of 16 subjects showed that the ITR of TRCA-metric learning algorithm was 6.1%±1.26%higher than that of TRCA algorithm in the optimal neighborhood parameters.The ITR of the online learning algorithm based on metric learning idea is increased by 4.19%±1.23%on average compared with the original online learning algorithm,and the improvement is more significant when the training data is small.