Nonnegative Matrix Factorization For EEG Feature Extraction Method

A motor imagery based Brain-Computer Interface(BCI) detects the electroencephalogram(EEG) with the power and phase changes over the scalp, and then determines the movements of Brain image(such as left/right hand movements). It is a system that analyzes the brain activities under specific conditions by using the computer or other electrical devices to transfers brain information, without depending on brain’s channels of peripheral nerves and muscles.Non-negative Matrix Factorization(NMF) is a new factor method for pattern recognition. It aims to find two non-negative matrices whose products can well approximate the original matrix. Motor imagery feature technique based on non-negative Factorization has good local characteristics and some sparsity. Furthermore, it selects meaningful time-frequency features of EEG automatically.Firstly, this paper described the time-frequency diagram of event-related spectral perturbation and analysed the mechanism of event-related desynchronization/synchronization phenomenon of the public data set of 2003 International BCI Algorithms Contests, then preprocessed on the left vs. right hand motor imagery EEG of single-trial and three electrodes, including: the choice of the channel, the selection of the EEG rhythms and the determination of the time-frequency range. The extracted feature vectors directly affect the classification results of the BCI, but the traditional feature extraction methods are always difficult to completely express the feature information contained in the EEG. Therefore, Non-negative Matrix Factorization(NMF) was applied to wavelet transformed EEG. We obtained the time-frequency representation of the EEG data, by filtering the characteristic frequency component of μ rhythm and β rhythm. We investigated the power and sparseness of each data vector to select Non-negative Matrix Factorization with Sparseness Constraints(NNFSC) seeks a time-frequency matrix in the decomposition. Then Independent Component Analysis was used to extract space domain. The method is a more complete expression of the time, frequency and spatial characteristics of EEG signals, and provides better feature vectors for classification.Then this paper used support vector machine method to recognize the classes of the feature vectors. In order to improve the performance of SVM classifier using radial basis function(RBF) kernel, this paper used genetic algorithm to set the value of the trade-off parameter C and the kernel parameter σ. Finally, comparing with the winner of BCI competition, the classification results improved, which verified the feasibility and effectivity of the designed method.