In order to solve the problem that invalid data affects the accuracy of EEG classification, a method of data screening is proposed. Based on brain computer interface (BCI) system, this paper presents an approach that using BP neural network to classify the EEG data generated by visual stimulation. The statistical characteristics of EEG signals corresponding to left and right motor imagery tasks are input to the BP neural network. First, the invalid data are eliminated by using the energy characteristics of β rhythm signal, and then classified by combining the mean value, standard deviation, energy spectrum, power spectrum, autocorrelation function and other features of μ rhythm signal. The using of β rhythm signal makes the characteristics more accurate and improves the accuracy of signal classification from 78. 25% to 84. 11%.