Classification of Subject-Independent Motor Imagery EEG Based on Relevance Vector Machine

Abstract

Brain-computer interface (BCI) is a novel technology that enables human to communicate directly with the outside world through brain consciousness. At present, there are problems such as complicated and time-consuming data collection (calibration process) and lack of sample set in the traditional BCI. Compared with the subject-dependent (SD) BCI, the subject independent (SI) BCI technology can shorten or not go through the calibration process and directly use the general model. It shortens the calibration process and meets the actual user’s requirements for the use of the BCI. In this paper, we design a method for classifying subject-independent motor imagery EEG signals based on relevance vector machine (RVM). By training the existing data of multiple subjects to obtain a general model, new subject can use the BCI system without the calibration process. This method uses the common spatial pattern (CSP) combined with the relevance vector machine classification algorithm, and uses a public data set (BCI competition IV-II-a) to verify its effectiveness. The offline experimental results show that for the classification of SI motor imagery EEG, the average classification accuracy of this method can reach 71.95%, and the highest accuracy can reach 96.14%, which proves that this method can effectively classify SI motor imagery EEG signals, and can achieve satisfactory results.