Adaptive Algorithms For EEG Motor Imagery-based Brain-Computer Interface System

The motor imagery-based brain-computer interface(MI-BCI)system decodes the neural response patterns by well-designed algorithms and converts them into computer commands to control external devices,which has broad application prospects in neural rehabilitation,functional prosthesis and cognitive mechanisms.Due to noise sensitivity and easy overfitting in small training set,a stable pattern recognition model requires substantial training data and a long calibration time.Additionally,the nonstationarities in EEG signals result in poor adaptability,and individual differences lead to poor transfer ability between subjects.To shorten the calibration time and improve the adaptability of the model,this thesis intends to train the feature extractor with the test data of the target subject or the training data from other subjects.Since the covariance matrix of EEG signal encodes the spatial discriminative information of the mental tasks,this thesis takes the covariance matrix as the feature descriptor and proposes a series of adaptive algorithms based on the decoding of covariance matrix for cross-session or cross-subject scenarios.The detailed contents of the thesis are as follows:(1)Considering the poor adaptability in cross-session,an adaptive common spatial patterns(ACSP)framework was proposed.ACSP realized a fast decomposition of the CSP filters by the recursive least squares(RLS),and a surpervised estimation of the intraclass covariance matrix with the proposed sample filter based on natural neighbor.ACSP algorithm not only alleviated the overfitting caused by insufficient training sets,but also improved the adaptability of the model.Satisfactory results had been achieved in both small training set and long duration online testing.(2)To tackle the overfitting and noise sensitivity of CSP model,two frameworks of geometry-aware CSP(ga CSP)were proposed by capitalizing on the geometric properties of the Riemannian manifold.a)The geometry-aware CSP with maximum between-class distance(ga CSP-B)generalized the joint diagonalization principle from Euclidean to Riemannian manifold,and performed a joint diagonalization on the Riemannian manifold while maximizing the between-class distances.b)The geometry-aware CSP with maximum within-class variance(ga CSP-W)formulated a dimensionality reduction framework from a high-dimensional Riemannian manifold to a low-dimensional Riemannian manifold while maximizing intraclass discriminative information.A series of experiments on two BCI competition datasets demonstrated the competitive results over state-of-the-art methods and confirmed the feasibility and effectiveness of proposed algorithms.(3)The domain adaptation is an effective technology to remedy the shortage of target data by leveraging rich labeled data from the sources.In this thesis,a domain adaptation based on Riemannian manifold embedding(e SPDA)was designed by combining the ga CSP dimensionality reduction concept and the domain adaptation method.Based on the principle of maximizing the distance between classes,the labeled source data was embedded into the more discriminative submanifold,and the principal characteristics of the unlabeled target data were preserved by the principle of maximizing variance within classes.Meanwhile,the domain adaptation method was integrated to minimize the distribution divergences between the source and target domains.A series of experiments conducted on two public datasets validated the validity of e SPDA.(4)Aiming at the feature transfer across subject,a kernel-based manifold domain adaptation(KMDA)algorithm was proposed.Instead of embedding covariance matrix into a low-dimensional Riemannian manifold,the KMDA algorithm mapped the Riemannian matrix to a high-dimensional Reproducing Kernel Hilbert Space through a kernel defined on the manifold.Then a subspace learning was performed by minimizing the conditional discrepancy between the sources and the target while preserving the target discriminative information.Additionally,an approach to convert the EEG trials into 2D frames(E-frames)was presented to further lower the dimension of covariance descriptors.Experiments on three EEG datasets demonstrated that KMDA outperforms several state-of-the-art domain adaptation methods in classification accuracy.KMDA showed potential in addressing subject dependence and shortening the calibration time of motor imagery-based brain–computer interfaces.