Research On Heterogeneous Label Space Migration Of EEG Signals Based On Riemannian Manifol

Brain-Computer Interface(BCI)technology is a human-computer interaction technology based on various neural signals,including Electroencephalogram(EEG).BCI system aims to establish interaction between the human brain and computers,and enables direct control of computers or other external devices by the human brain.By capturing,processing,classifying,converting,and providing feedback on EEG signals,it achieves high-speed,accurate,and direct interaction between humans and computers.Brain-Computer Interface offers greater autonomy and more effective control for people with disabilities,as well as more efficient and convenient control for ordinary individuals.However,Brain-Computer Interface requires the collection of sufficient EEG signals to train a classification model.When BCI systems operate in larger dimensions,this can place a significant burden on data collection,potentially reducing the practicality of brain-computer interfaces.In this paper,how to utilize transfer learning to reduce the data collection burden on subjects is explored in a brain-computer interface system based on speech imagery and motor imagery.Specifically,the concept of transfer learning is introduced in this paper.Labeled EEG samples from other subjects are used to train the classification model of the current subject so that a well-performing classifier is obtained.Differing from traditional transfer learning algorithms,Multi-band Data Stitching with Label Alignment and Tangent Space Mapping(MDSLATSM)heterogeneous label space transfer learning algorithm is designed based on the sequential characteristics of the implemented experimental paradigm.By acquiring a large number of artificially labeled EEG samples from the current subject,the amount of data that needs to be collected is reduced by the algorithm.Thus,the collection burden on subjects is alleviated.The main research components of this paper are as follows:(1)Based on an in-depth study of two different imagery tasks,language imagery and motor imagery,a novel multimodal experimental paradigm is designed.The experimental paradigm aims to enhance the set of operational instructions of the brain-computer interface system so that further improve its performance.Specifically,a strategy of sequential encoding is adopted in this paper.Based on the two single kinds of speech imagery and motor imagery tasks,sequential encoding on the two single kinds of tasks is performed individually,thus achieving the goal of obtaining multimodal EEG signals.(2)Based on the proposed experimental paradigm,after fully analyzing the EEG signals of this experimental paradigm having cross-label transferability characteristics,a data stitching algorithm is proposed in this paper.By this algorithm,the labeled data are added artificially to reduce the acquisition burden of the subjects.Then the acquired multimodal EEG signals are processed by the label alignment algorithm.Therefore,the distribution differences are reduced significantly between the acquired artificially labeled training data and the existing labeled data.(3)For the experimental paradigm of sequential coding,a Multi-band Data Stitching with Label Alignment and Tangent Space Mapping(MDSLATSM)transfer learning algorithm is proposed.Firstly,the EEG data with more information is obtained by a multi-band filtering algorithm.Then,the source domain data is stitched into artificial signals by a data stitching algorithm.Those artificial signals share the same sequential features as the target domain data so that builds a bridge between the source and target domain.The covariance matrix of the two domains is aligned by label alignment to make the data distribution of the two domains closer.After the data is mapped into the tangent space,features are extracted from the Riemann manifold.Finally,a high classification accuracy is derived through feature selection and classification.The EEG signals of 16 subjects are used as the dataset for heterogeneous transfer learning experiments.For the heterogeneous transfer learning scenarios of cross-label,the average classification accuracy is 78.28%.Meanwhile,MDSLATSM is also tested cross-subject,with an average classification accuracy of 64.01%.The experimental results show that compared to traditional EEG signals recognition algorithms,classification accuracy and generalization performance are improved.The data collection burden on subjects is also reduced by the MDSLATSM algorithm.Therefore,the development of BCI technology is advanced.