Classification of Four Categories of EEG Signals Based on Relevance Vector Machine

EEG classification is an important signal acquisition equipment of brain-interface research and application. In this paper, an improved relevance vector machine (RVM) is proposed to classify four-class motor imagery EEG signals. The original EEG signals are processed by 3-24Hz band-pass filter. Thereafter, EEG feature vectors are extracted from the band-pass filtered EEG signals with one versus one common spatial patterns (OVO-CSP). Then, the improved RVM algorithm with kernel function, which combines Gaussian kernel function and Cauchy kernel function, is applied to classify the EEG signals. A public dataset (BCI Competition IV-II-a) is employed to verify the proposed improved kernel function method. Five-fold cross-validation was used to ensure that the classification of the experiment is more credible. The classification results of the combined kernel function are compared with results of the Gaussian kernel function and the Cauchy kernel function. The experimental results show that the highest classification accuracy of the proposed kernel function and the single kernel function in the first data set are 64.40% and 60.60%, respectively, and in the second data sets are 67.58% and 63.33%. The average classification accuracy of the proposed kernel function is 4%-4.19% higher than that of the single kernel function. These results show that the proposed method is advantageous in the classification of four-class motor imagery.