The Research Of EEG Signals Classification Based On Ensemble Extreme Learning Machine

Electroencephalogram (EEG) is the comprehensive reflection of the state of the electrical activity of brain tissue and brain function. It records the physiology electrical activity of our brain through the electrodes distributing on the brain, and it is a kind of highly complex random signal that contains a variety of information. For a long time, EEG signals classification is used to identify the patient’s diseases (such as epilepsy, schizophrenia, etc.), which achieves good results. However, the traditional method by artificial reading has not satisfy the needs of EEG signals analysis at the present times. Therefore, automatic EEG signals classification with intelligent methods such as machine learning is becoming the focus of the study.The process of automatic EEG signals classification includes two stages:feature extraction and classification. In the phase of feature extraction, to solve the problem that traditional methods can only obtain the single property of signals, this paper proposes a hybrid feature extraction method based on wavelet packet and sample entropy. The method uses wavelet packet decomposition to extract the linear characteristics of signal and gets the nonlinear characteristics with sample entropy, which can describe both linear and non-linear dual characteristics of EEG signals better. In the phase of classification, ensemble learning is adopted in this paper in order to solve the problem of unstable predicted results and poor generalization ability, and two improved ensemble extreme learning machine methods based on different ensemble strategies are proposed. Ensemble extreme learning machine based on linear discriminant analysis uses parallel ensemble strategy. Linear discriminant analysis is used to get a subset with the larger diversities between each other and increase the differences between each learning machine. Adaboost extreme learning machine based on mutual information takes advantage of serial ensemble strategy. The single extreme learning machine pays more attentions on the error classification samples. Mutual information variable selection is embedded and the performance of strong learning machine is regarded as evaluation index to optimize the input variables and network model. Experiment results show that the proposed feature extraction method can get comprehensive information, which is advantageous to the classifier and the proposed two ensemble classification methods can effectively improve classification accuracy, and have better generalization performance.