A Multi-Domain Feature Extraction Method with Improved LMD

Motor Imagerary EEG (MI-EEG) signals have non-stationary and nonlinear characteristics, it is necessary to extract and fuse the features in multi-domain. Local Mean Decomposition (LMD) algorithm has been applied to feature extraction of MI-EEG. However, it is limited due to mode mixing and endpoint effect. In this paper, the traditional LMD algorithm is improved for feature extraction and fusion in multi-domain. We use pairwise complementary white noise and extremum extension to improve LMD algorithm, and denoted as CEELMD. Then, use CEELMD to decompose each channel of EEG signal to obtain multiple product function (PF) components. According to the frequency band of motor imagery to choose PF components, and their multi-scale fuzzy entropies (MFE) are calculated as nonlinear features, the CSP algorithm is used to extract the spatial features. Furthermore, the features are fused in serial and evaluated by the random forest algorithm. In this paper, the experiments are conducted on the BCI IV 2a competition data set and the average classification accuracy reaches 75.27%. The experimental results show that CEELMD can effectively improve mode mixing and endpoint effect, it is helpful to feature extraction for distinguishing four types of motor imagery EEG signals.