Optimal Frequency Band Selection Based on Filter Banks and Wavelet Packet Decomposition in Multi-class Brain–Computer Interfaces

To address the problem of low system recognition rate in multi-class brain–computer interface (BCI), many ideas have been proposed. Among them frequency optimization is a good option because the performance of BCI systems depends largely on the operational frequency band of electroencephalography (EEG) signals. In this chapter, feature selection method was utilized to select frequency bands for classification in multi-class BCI systems. The procedures of the proposed algorithm are as follows: Firstly, all single-trial EEG signals are divided into some subband signals by two different methods. One is wavelet packet decomposition and the other is filter bank based on coefficient decimation. Secondly, a multi-class common spatial pattern algorithm based on approximate joint diagonalization is used to extract features of EEG signals on each subband. Finally, optimal subband features are selected for classification. In offline analysis, the proposed method yielded relatively better cross-validated classification accuracies.