Using Convolution Neural Networks Pattern for Classification of Motor Imagery in BCI System

The Electroencephalography (EEG) based Brain-computer interfaces (BCI) enable humans to control external devices through extracts informative features from brain signals and convert these features into control commands. Deep learning methods have been the advanced classification algorithms used in various applications. In this paper, the informative features of EEG signals are obtained using the filter-bank common spatial pattern (FBCSP), then the selected features which are prepared using the mutual information method are fed to the classifiers as input. Convolution neural network (CNN), Naive Bayesian (NB), multiple support vector machines (SVM) and linear discriminant analysis (LDA) algorithms are used to classify EEG signals into left and right hand motor imagery (MI) across nine subjects. Our framework has been tested on BCI competition IV-2a 4-class dataset. The results are shown that the CNN classifier has yielded the best average classification accuracy, with 99.77% as compared to other classification methods. The experimental results represent that our proposed method can obtain more refined control in the BCI applications such as controlling robot arm movement.