Research On EEG Classification Algorithm Based On Transfe Learning

Brain computer interface is a technology that directly establishes information interaction between the brain and external devices,and achieves mutual influence and control.It can decode brain activity patterns from EEG signals recorded during specific psychological tasks,converting neural responses into control instructions.The brain computer interface paradigm based on motor imagery is one of the most concerned paradigms at present.Motor imagery has been widely applied in medical rehabilitation,and stroke patients can achieve rehabilitation training through motor imagery to improve their motor function and ultimately restore control over body parts.Due to the high cost of EEG signal acquisition,the number of samples per subject is usually small,and these signals have non-stationary and nonlinear characteristics,resulting in individual differences in signal distribution among different subjects.Therefore,it is not possible to directly use cross subject EEG signals to construct robust classification models.Therefore,this paper hopes to reduce the individual difference of EEG signal distribution through transfer learning,and build a highly robust classification model on the EEG signals of all cross subjects.The specific content of this article is as follows:(1)A motion imagination classification algorithm based on parameter migration is proposed,which uses the EEGNet network and channel attention mechanism,and introduces the fine-tuning technology in transfer learning to improve the pre training model.According to the classification accuracy of public and self collected datasets,the method proposed in this chapter can effectively improve the accuracy of motion imagination classification.(2)A motion imagination classification algorithm based on instance transfer is proposed.The kernel mean matching algorithm is used to obtain the weight sample matrix between the source domain and target domain subjects,and this matrix is used as the initial weight matrix of Tr Ada Boost.Finally,the strong classifier trained by Tr Ada Boost is used for classification.According to the classification accuracy of both publicly available and self collected datasets,the method proposed in this chapter can effectively utilize source domain data to improve the classification accuracy of the target domain.(3)A motion imagery classification algorithm based on heterogeneous migration is proposed.The Heterogeneous Label Spaces Alignment algorithm is used to align the source user category to the target user category,extract the covariance matrix of the sample,calculate the tangent space vector as the input,and finally use the joint distribution adaptive method to reduce the distance between the joint probability distribution of the source domain and the target domain,and use the naive Bayes classifier to classify the data.According to the classification accuracy of public datasets,this method can effectively reduce the distribution differences between the source domain and the target domain,and improve the classification accuracy of the target domain by utilizing data with different category spaces.(4)A single channel EEG artifact removal algorithm is proposed to address the limitations of multi-channel EEG devices in the medical field,laying a foundation for the application of single channel EEG devices in the medical field and the analysis and processing of EEG signals in stroke patients.