Research Of Feature Selection And Classification Methods In Brain-Computer Interface

Brain-computer interface (BCI) is a direct communication and control channel between human brain and computer or other electronic device. It does not depend on the brain’s normal output pathways of peripheral nerves and muscles. The BCI is a novel kind of human computer interface and recently it is an active topic in brain function research. BCI technology can help improve the quality of life and restore function for people with severe motor disabilities. BCI research is a multidisciplinary field and has been a hot focus in brain science, rehabilitation engineering, biomedical engineering and human computer interaction.Feature selection and classification methods of EEG signals are two key points of BCI technology. Feature extraction methods of EEG signals using in BCI systems are discussed in this paper, mainly including modern power spectral density (PSD) and wavelet analysis. Then classification methods used in BCI systems are investigated thoroughly, mainly including Fisher linear discriminant analysis (LDA), neural network (learning vector quantization neural network and probabilistic neural network) and support vector machine. In this paper, these feature selection and classification methods are used in the off-line analysis of some typical BCI dataset, and an asynchronous BCI system based on Alpha wave is built. Main contributions of this paper include:(1) A typical BCI system using slow cortical potential (SCP) is off-line analysed. Features are got from the SCP, and LDA is used for classification. This method has got good classification accuracy, fast speed and good stability. The result of the experiment shows that training can give much help for the subject to master skill and improve classification accuracy. At the same time, a simulation model based on Simulink is built, which is very flexible and can be widely used in SCP-based BCI system.(2) Relative wavelet energy (RWE) used for feature selection of EEG is proposed. This method is utilized in BCI system using imaged left or right hand movement. Classification accuracy and mutual information (MI) are used for evaluation criteria of BCI system. LDA and SVM are respectively utilized for classification and compared with each other. The results of the experiment show that RWE is a good feature selection method, comparing with AAR. And MI is more reliable than classification accuracy.(3) PSD using multiple signal classification (MUSIC) method for feature selection of EEG is proposed, and learning vector quantization (LVQ) neural network is used for classification. Comparing with AR model, the PSD using MUSIC method has better separability and stability. The experiment result shows that this method provides a new way for BCI research.4) For BCI system using electrocorticography (ECoG) signals, feature selection method using RWE is proposed. Principal components analysis (PCA) is used to reduce the dimension of the feature vector which gets from RWE. Probabilistic neural network (PNN) is used for classification, and the influence of the spread of radial basis functions to the classification accuracy is investigated. The result of the experiment shows that this algorithm has got significant improvement on classification accuracy.5) An asynchronous BCI system platform based on Alpha wave is built, and it is real-time analysed. The basic theory of this BCI system is alpha wave-block phenomenon. It mainly includes:real-time EEG acquisition, preprocessing, feature selection and classification. An asynchronous working mode which is suitable for BCI system is proposed. The subject can decide freely when he or she wishes to run the BCI system and chooses anyone of four commands as output. It is a more natural human computer interaction method. This system has got high classification accuracy, good stability and practicability.