Research On Motor Imagery EEG Decoding Method Based On Sparse Optimization

The brain-computer interface(BCI)system based on motor imagery is a very important link in rehabilitation medicine and rehabilitation engineering training.However,motor imagery electroencephalography(EEG)have low signal-to-noise ratio and strong randomness,especially the outstanding individual differences and non-stationarity,resulting in low and unstable decoding accuracy of the motor imagery EEG,which directly affects the effectiveness of rehabilitation functions.In order to further improve the performance of motor imagery EEG decoding,combining the characteristics of EEG signal and physiological prior information,the motor imagery EEG decoding method has been studied systematically in this paper based on the theory and method of sparse optimization.The specific research work is as follows:First,in view of the channel selection and classification of EEG signal and the limitations of cross-validation,a motor imagery EEG signal classification model based on group sparse Bayesian logistic regression was proposed.Under the framework of Bayesian learning,group sparse modeling of EEG channels was carried out by introducing grouped automatic relevance determination(GARD)priori.The new model can perform channel selection and classification at the same time,and model parameters can be automatically estimated from the training data,avoiding the tedious and time-consuming cross-validation process.The experimental results show that the proposed method achieves better classification accuracy and fewer channels,and the selected channels are more consistent with neurophysiology.Second,in order to improve the accuracy,adaptability and physiological interpretability of the EEG signal classification model,a new motor imagery EEG classification model was proposed,that is,the fused group least absolute shrinkage and selection operator(LASSO).On the one hand,the features were grouped according to the channel attributes,and the features of the same channel were assigned the same weights to perform group sparse modeling;on the other hand,the total variation norm was used to constrain the weights of adjacent channels to be the same or similar,so as to achieve spatial smooth modeling.The experimental results show that,compared with the existing sparse optimization methods,the proposed method has better classification accuracy and physiological interpretability;compared with the existing spatial filtering methods,the global spatial smoothness is realized by a data-driven approach in the proposed method,so the decoding model is more adaptive.Third,in order to solve the problem of large computation and time-consuming in the existing common spatial pattern(CSP)feature extraction and feature selection methods,three new CSP feature extraction methods and a non-convex logarithm sparse feature selection method were proposed.In the aspect of feature extraction,the EEG signals were spatially filtered using CSP,and then the frequency domain information of CSP was compensated by three methods including discrete wavelet transform(DWT),wavelet packet decomposition(WPD)and filter bank(FB).In the aspect of feature selection,in order to solve the bias problem of LASSO,a non-convex regularization model based on log function was proposed to select more discriminative spatial-frequency features.In addition,in order to further optimize feature selection and enhance the robustness of the classification model,an ensemble learning method for secondary feature selection and ensemble model construction was proposed.The experimental results show that the proposed method has achieved better classification performance and lower feature extraction time than other methods,and the performance of feature selection is also better than existing methods.Fourth,in view of the noise sensitivity and over-fitting problem of the CSP method,the joint selection problem of time-space-frequency and the non-stationarity problem of the EEG signal,a motor imagery EEG decoding method based on LASSO sparse feature selection and Tikhonov regularization CSP(TRCSP)ensemble was proposed.A unified algorithm framework was established by the organic combination of CSP regularization,time-frequency joint optimization and ensemble learning.Under the framework,CSP noise sensitivity and overfitting problems,time-space-frequency joint selection problems and non-stationary problems of EEG signals were solved simultaneously.The effectiveness of the proposed method is verified by the experimental results of a large amount of data.Compared with the existing CSP time-frequency optimization method,the proposed method has a small amount of calculation,low model complexity,and has better robustness and stability.In this paper,some new theories and methods are proposed to solve the decoding problem of motor imagery EEG,and some results have been obtained.The proposed method has certain reference value and promotion effect for future BCI system development.