Semi-supervised Broad Learning System Algorithm For EEG Classification

BCI is a system that controls external equipment through brain thoughts.The development of BCI system has practical applications in many fields,such as wheelchair control,AI(Artificial Intelligence)intelligence,virtual reality games,etc.In the field of biomedicine,the BCI system mainly provides technical support for patients with mobility impairments to restore their exercise capabilities.In BCI system,electroencephalogram is one of the most used EEG signals,which has the function of reflecting brain physiology,thoughts and emotions.How to effectively and quickly classify EEG signals to judge brain thought is one of the keys to BCI technology.Therefore,it is important to find a classification algorithm with good generalization and fast speed.This paper carried out the research of EEG signal classification based on Broad Learning System(BLS).The main work is as follows:(1)In view of the original broad learning using only labeled samples to be applied to the supervised field,Graph Semi-Supervised Broad Learning System(GSS-BLS)is proposed.First,extract the Common Spatial Pattern(CSP)feature of the original EEG signal,then use the graph label diffusion method to label the unlabeled samples,and add it to the semi-supervised broad learning,and finally perform the test datasets.Experiments have confirmed that GSS-BLS has good performance on 3 EEG datasets,which is better than the supervised BLS algorithm,and compared with other classification algorithms,it also shows that GSS-BLS has an ideal effect on EEG signal classification.(2)Aiming at the safety problem of semi-supervised learning,a Transfer Semi-Supervised Broad Learning System(TSS-BLS)algorithm is proposed.Introduce the distribution adaptation method in transfer learning and manifold regularization term to broad learning,and transfer labeled samples to unlabeled samples in order to reduce the security problem of unlabeled samples.At the same time,the manifold regularization term uses unlabeled sample information to further improve TSS-BLS classification performance.The TSS-BLS algorithm is tested on three EEG datasets and compared with 7 methods.The experiment confirms that TSS-BLS is more suitable for EEG signal classification than other algorithms.In addition,this paper conducts classification experiments on some UCI public datasets,and the results further prove the effectiveness of TSS-BLS.