| Brain computer interface (BCI) is a communicational and control channel whichis established directly between human and computer or other electronics deviceswithout the participation of peripheral nerve or muscles，and through translating brainsignals into control commands to achieve communication and control with outsideworld. With its important theoretical value and actual application prospect, thetechnology has become one of the hotspots in biomedical signal processing. Theresearch of motor imagery (MI) EEG is an important branch in BCI.This study is based on the requirements of the National Natural ScienceFoundation (61201302). Starting from the background and significance of theresearch, this thesis describes the characteristics of EEG, and analyzes the method ofMI EEG preprocessing, feature extraction and pattern classification. Furthermore thethesis introduces the collection device and acquisition scheme of EEG, translates thespecific MI tasks into corresponding control commands which is put into the electricwheelchair to finish the particular movements. In this thesis, the following researchwork has been done with considerable research results.(1) Pre-processing: The algorithm of optimizated generalized weightedestimators (OGWE) is used to preprocess EEG, which can eliminate the signals thathave no connection with MI EEG and can enhance the signal-to-noise ratio, andprovide a good basis for feature extraction and pattern classification.(2) Feature extraction: From the perspective of conventional MI activatingpartial brain areas, as there are many irrelevant frequency EEG signals in MI EEGand the feature extraction method of common spatial pattern is the lack of frequencyinformation processing, feature extraction method combining dual-tree complexwavelet transform (DTCWT) with common spatial pattern (CSP) is presented. FirstlyEEG signals of specific channels are selected and dual-tree complex wavelet is takento make multi-scale decomposition to get EEG signals of appropriate bands. And thenall these are combined to be input to the spatial filter of common spatial pattern to geta needed feature vector. From the perspective of a complex functional brain network(FBN), a novel method is proposed based on adjacent matrix decomposition offunctional brain network in this thesis. Firstly, multi-channel motor imagery EEG signals are used to construct the FBN. Secondly, the corresponding adjacent matrix isused to make singular value decomposition to get singular value characteristic vector.And then the characteristic parameters of EEG are defined based on singular valuecharacteristic vector. Finally feature vector is consisted of these parameters.(3) Pattern classification: In order to enhance classification accuracy and speedin the BCI system, a spindle dynamic kernel clustering method based on deepautoencoders (DA) dimension reduction is presented in this thesis. Firstly, the DAalgorithm is used to reduce the eigenvector into two dimensions in order to lower thecorrelation and computational complexity among feature vector. And then the spindledynamic kernel clustering is taken to make classification. In addition, as problemssuch as small sample estimation, nonlinear, non-stationary signal classification can besolved by the support vector machine (SVM), a classifier based on multi-classmultiple kernel learning (MKL) SVM is designed, by which complex data can bemapped into high-dimensional space better, and the classification accuracy can beimproved while the number of support vectors can be reduced.(4) Experiments of controlling an electric wheelchair: Firstly, four classes of MIexperimental paradigms are designed, and then the OGWE method is used for blindsource separation. Secondly, DTCWT-CSP is taken to extract the feature vectors ofEEG. MKL-SVM is then adapted for the classification and recognition of thesevectors. Finally, the results are turned into control commands to control the motionsof the electric wheelchair. The average accuracies of three subjects are66.78%,76.58%and72.53%, respectively. |
PDF Full Text Request