Neural Network Based BCI by Using Orthogonal Components of Multi-channel Brain Waves and Generalization

FFT and Multilayer neural networks (MLNN) have been applied to ‘Brain Computer Interface’ (BCI). In this paper, in order to extract features of mental tasks, individual feature of brain waves of each channel is emphasized. Since the brain wave in some interval can be regarded as a vector, Gram-Schmidt orthogonalization is applied for this purpose. There exists degree of freedom in the channel order to be orthogonalized. Effect of the channel order on classification accuracy is investigated. Next, two channel orders are used for generating the MLNN input data. Two kinds of methods using a single NN and double NNs are examined. Furthermore, a generalization method, adding small random numbers to the MLNN input data, is applied. Simulations are carried out by using the brain waves, available from the Colorado State University website. By using the orthogonal components, a correct classification rate Pc can be improved from 70% to 78%, an incorrect classification rate Pe can be suppressed from 10% to 8%. As a result, a rate Rc = Pc/(Pc+Pe) can be improved from 0.875 to 0.907. When two different channel orders are used, Pe can be drastically suppressed from 10% to 2%, and Rc can be improved up to 0.973. The generalization method is useful especially for using a sigle channel order. Pc can be increased up to 84 ∼ 88% and Pe can be suppressed down to 2 ∼ 4%, resulting in Rc = 0.957 ∼ 0.977.