Research On Decoding Algorithm Of Motor Imagery EEG Signals Based On Covariance Features

The motor imagery EEG signal is the EEG signal generated by the subject spontaneously imagining the movement of the limbs.It can complete the corresponding task without external stimulation.The interaction method is more friendly and natural,and it is closer to conscious control.Brain-computer interface systems based on motor imagery EEG signals have great research and application value in medical rehabilitation,human-computer interaction and other fields.The original motor imagery EEG signal has a very low signal-to-noise ratio,and the useful information contained in it is extremely weak and easily affected by the environment,which makes motor imagery EEG decoding very difficult.Motor imagery EEG decoding methods based on covariance features have made great progress in recent years,achieving state-of-theart performance on multiple datasets.Although the decoding method based on covariance features is very efficient,the Riemannian geometric tools used to deal with covariance features involve a large number of matrix decomposition operations.When the dimension of covariance features increases,the computational load of the algorithm will increase sharply;Motor imagery EEG signals also have strong individual differences,and the data between different subjects is quite different.It is impossible to use the existing samples to correct the current user’s braincomputer interface system,which restricts the application of the brain-computer interface system.This research focuses on the above two issues,and the main contents include:1.The feature dimension of the covariance matrix is high,which can easily lead to high computational complexity and overfitting.Therefore,for high-dimensional raw covariance features,we propose an EEG classification framework based on Riemannian submanifold feature fusion.This method aims to reduce the dimension of the Riemannian manifold while enhancing the separability of features.The method constructs a Riemannian manifold on three frequency bands related to motor imagery,and then uses the improved Riemann CSP method to solve the bilinear dimensionality reduction matrix in the three frequency bands respectively to construct the Riemann submanifold.Finally,in the Riemann subflow On the tangent space of the shape,the mutual information score is used to select and fuse features to obtain separable features efficiently.This method not only reduces the computational complexity of the algorithm,but also enhances the separability of features,and outperforms other comparable algorithms on BCI competition Ⅳ2a data and Tsinghua University data.2.In order to solve the problem of individual differences in motor imagery EEG signals,we propose a filter bank Riemannian Tangent Space(FBTSCORAL)method to reduce the distribution of different subjects’ data.differences,and improve the accuracy of transfer across time periods and individuals.The method constructs multiple Riemannian manifolds on the filter bank and splices the features on multiple Riemannian spaces as sample features,and then uses the Correlation Alignment(CORAL)method to align the features of the source and target domains.Finally,it is used for cross-period transfer,one-to-one crossindividual transfer and many-to-one cross-individual transfer.The method is experimented on the BCI competition Ⅳ2a dataset and outperforms the performance of other comparable methods.