Research On Feature Exaction And Classification Methods For EEG

Brain-Computer Interfaces(BCI) provides a direct communication tool for brain and external, which has been extensively studied by many scholars. Electroencephalogram(EEG) has become the main data source of BCI research due to its lower cost. Therefore, EEG plays an important role in the life and scientific research. For BCI studies, nonlinear characteristic and instability of EEG signals leads to enormous challenges and difficulties. Extracting effective feature classification from the original signal plays a vital role.Feature extraction and classification are two main jobs of EEG signal process. Feature extraction can adopt one method, such as autoregressive(AR), sample entropy, approximate entropy, wavelet packet decomposition, fast Fourier transform, Hilbert transform; it also can adopt a combination of two methods, such as empirical mode decomposition and combined with intrinsic mode function, phase locking value with Hilbert transform, and wavelet with entropy and so on. This paper aims to study feature extraction and classification methods. The main work of the paper is as follows.1. A single method for extracting feature. This paper makes a contrast experiment for a single feature extraction method. In the proposed method, feature extraction uses sample entropy and autoregression(AR) model, respectively. The classifier is support vector machine and its parmenters are optimized. Experimental results verified by five mental tasks show that the processing time of two feature extraction methods is short. Compared to the sample entropy tedious selection parameters, autoregression classification model achieves better results.2. Extract feature combined two methods. Aiming at non-stable and non-linear characteristics of EEG signals, this paper proposes a novel EEG classification method to improve the classification accuracy and processing time. For feature extraction method, the proposed method combines wavelet package decomposition and sample entropy, and presents a feature exaction method of sample entropy based on wavelet package decomposition. For the selection of classifier, the proposed method employs extreme learning machine. The paper has also done a lot of comparative experiments to verify the proposed method, based on approximate entropy of wavelet packet decomposition as feature extraction method, support vector machine as classifier; and sample entropy of wavelet packet decomposition as feature extraction method, extreme learning machine as classifier. Experiments are performed in epileptic EEG data and five mental tasks, respectively. Experimental results show that the proposed combination strategy of sample entropy and extreme learning machine has shown great performance, which obtains good classification accuracy and low training time.