An Approach Based on BP Neural Network and Optimal F Features for Classifying Single-Trial EEG

In order to increase the classification accuracy of single-trial classification of electroencephalogram (EEG), an approach of recognition of mental task based on BP neural network and optimal features is presented. According to the space distribution of EEG electrodes on the brain and the frequency characteristics of EEG and the noises, the low-pass filter and the spatial filter are adopted for obtaining the useful EEG components. The average value of each channel EEG signals are calculated based on bereitschaftspotential. The power spectral density of multi-channel EEG signals are analyzed based on autoregressive model. The electrodes are selected according to the obvious difference of the bereitschaftspotential and and the energy features in frequency domain between the two mental tasks on all the 28 channels. The optimal features both in time and frequency domain are obtained by using the sliding window with adjustable window width and separability measure. Then the average values of each channel EEG signals from -100 to -50ms and from -50 to 0ms are calculated as the features in time domain. The frequency bands from 7Hz to 12Hz and 17Hz to 24Hz of multi-channel EEG signals are chosen for calculating the energy features in frequency domain. Furthermore, the BP neural network with 3 layers is used to recognize the mental tasks. The proposed approach has been applied to the data set IV of "BCI Competition of 2003". The classification accuracy is up to 84% on the test set. The results show that it is a promising approach for classifying single-trial EEG in BCI system.