Research On Motor Imagery Eeg Feature Extraction Method In Time-frequency-space Domain

The brain–computer interface (BCI) is a direct communication pathway betweenhuman brain and external device. The primary goal of BCI is to help severely disabledpeople to communicate with computers or control electronic devices through theirthoughts. In this paper, based on the motor imagery EEG of BCI, we did in-depth researchon several key techniques in the process of BCI development, such as feature extractionmethod, feature classification method and so on. And then we compared the method ofthis paper with the traditional methods.Firstly, this paper described the mechanism of event-related desynchronization /synchronization phenomenon, and then did preprocess on the BCI motor imagery EEGdata, including determining the channel, rhythms of EEG and the time-frequency range.The extracted feature vectors affect the classification results of BCI directly. Thetraditional feature extraction methods are always difficult to express the featureinformation contained in the EEG completely. Therefore, based on the analysis of thetraditional feature extraction method, such as wavelet transform or independentcomponent analysis method, a new method in time-frequency-space domain was designed.In this paper, Independent Component Analysis (ICA) and wavelet transform were used toextract the temporal, spectral and spatial features from the original EEG signals.Then support vector machine method was used to recognize the classes of the featurevectors. For SVM using radial basis function (RBF) kernel, the value of the trade-offparameter C and the kernel parameterσwere always depended on setting the range ofappropriate parameters set in advance. So this paper used genetic algorithm improved theSVM classifier.Finally，as the classification results showed, the numbers of features are 14 and themax classification accuracy is as high as 96.4%. Compared with the winner of competition,the result of this paper is increased. Meanwhile，the result is 3.4% higher than theaccuracy of tradition method. So the proposed method represented better classificationperformance and the result verifies the feasibility and effectivity of the designed method.