EEG time-frequency-spatial feature selection method based on non-convex log model

Aiming at the time-frequency-spatial feature selection problem of motor imagery electroencephalogram (EEG), a sparse feature selection method based on non-convex log model is proposed, which is called the LOG method. First, the original EEG signal is decomposed into multiple time-frequency segmentations. Second, the common spatial pattern (CSP) is used for feature extraction on each time-frequency segmentation to obtain a time-frequency-spatial feature set. Finally, a non-convex sparse optimization model based on the log function is proposed for feature selection and classification, which can effectively alleviate the biased estimation problem of L₁ norm regularization. To verify the effectiveness of the proposed method, experiments are conducted using three publicly available motor imagery EEG datasets. Compared with the existing convex sparse optimization model, the non-convex log model achieves the highest average classification accuracy of 82.5%. The experimental results show that the LOG method not only achieves better classification accuracy, but also has better stability and robustness.