A Dual-Size Convolutional Kernel CNN-Based Approach to EEG Signal Classification

Abstract

AbstractIn this paper, an algorithm based on a combination of Riemannian space and convolutional neural network is proposed for the feature extraction as well as classification of motor imagery EEG signals with four classifications (left hand, right hand, foot and tongue). In terms of motor imagery EEG signal processing and feature extraction, the covariance matrix of the symmetric positive definite matrix is chosen as the descriptor to reasonably transform the EEG signal described in the spatio-temporal domain from Euclidean space to Riemannian space, followed by a low-dimensional vector estimation of the n-dimensional symmetric positive definite matrix using a specific feature mapping to obtain a low-dimensional and efficient EEG signal feature set. The convolution operation is improved in the classifier by adding a double-size convolution kernel to capture the motor imagery EEG data features more comprehensively. Experimental testing of the method using the publicly available dataset BCI2008IV-2a competition number competition dataset shows that the algorithm can effectively identify EEG signals with high classification accuracy and good robustness.KeywordsMotor imagery EEG signalBrain-computer interfaceRiemannian spacesCovariance matrixConvolutional neural network