A reduced rank approach for covariance matrix estimation in EEG signal classification

Common Spatial Pattern (CSP) methods are widely used to extract the brain activity for brain machine interfacing (BMI) based on electroencephalogram (EEG). For each mental task, CSP methods estimate a covariance matrix of EEG signals and adopt the uniform average of the sample covariance matrices over trials. However, the uniform average is sensitive to outliers caused by e.g. unrelated brain activity. In this paper, we propose an improvement of the estimated covariance matrix utilized in CSP methods by reducing the influence of the outliers as well as guaranteeing positive definiteness. More precisely, our estimation is the projection of the uniform average onto the intersection of two convex sets: the first set is a special reduced dimensional subspace which alleviates the influence of the outliers; the second is the positive definite cone. A numerical experiment supports the effectiveness of the proposed technique.