Common spatial pattern (CSP) is a very popular method for spatial filtering to extract the features from electroencephalogram (EEG) signals, but it may cause serious over-fitting issue. In this paper, after the extraction and recognition of feature, we present a new way in which the recognition results are fused to overcome the over-fitting and improve recognition accuracy. And then a new framework for EEG recognition is proposed by using CSP to extract features from EEG signals, using linear discriminant analysis (LDA) classifiers to identify the user's mental state from such features, and using Choquet fuzzy integral to fuse classifiers results. Brain-computer interface (BCI) competition 2005 data setsⅣa was used to validate the framework. The results demonstrated that it effectively improved recognition and to some extent overcome the over-fitting problem of CSP. It showed the effectiveness of this framework for dealing with EEG.
Signal classification is a key of brain-computer interface (BCI). In this paper, we present a new method for classifying the electroencephalogram (EEG) signals of which the features are heterogeneous. This method is called wrapped elastic net feature selection and classification. Firstly, we used the joint application of time-domain statistic, power spectral density (PSD), common spatial pattern (CSP) and autoregressive (AR) model to extract high-dimensional fused features of the preprocessed EEG signals. Then we used the wrapped method for feature selection. We fitted the logistic regression model penalized with elastic net on the training data, and obtained the parameter estimation by coordinate descent method. Then we selected best feature subset by using 10-fold cross-validation. Finally, we classified the test sample using the trained model. Data used in the experiment were the EEG data from international BCI Competition Ⅳ. The results showed that the method proposed was suitable for fused feature selection with high-dimension. For identifying EEG signals, it is more effective and faster, and can single out a more relevant subset to obtain a relatively simple model. The average test accuracy reached 81.78%.