Ensemble learning method based on temporal, spatial features with multi-scale filter banks for motor imagery EEG classification

Abstract

Decoding motion intention from electroencephalogram (EEG) is a key part of the motor imagery-based brain-computer interface (MI-BCI). To help the disabled with neuromuscular disabilities to restore motor function through BCI, it is necessary to build an efficient and stable classification algorithm to decode the motor intention contained in the EEG signal. However, EEG signals are non-stationary and vary greatly between individuals. In this work, we propose an ensemble learning method based on temporal, spatial features and multi-scale filter banks, called TSMFBEL, which aims to design an ensemble classifier model with strong generalization capabilities for MI EEG classification. To obtain diverse ensemble classifiers, the original EEG data are divided into subsets with sample diversity by bootstrap sampling method, and then decomposed into time–frequency subsets with time–frequency distribution diversity by multi-scale filter banks method. For each time–frequency subset, features with domain diversity are extracted from temporal domain, spatial domain and temporal-spatial domain, and heterogeneous classifiers with diversity are trained based on each set of features. To obtain the optimal decision, we describe the ensemble strategy as a minimum classification error optimization problem, and propose an ensemble classifier weight optimization method based on the L2-norm, and finally integrate the decision of the ensemble classifier by weighted fusion. The proposed method was evaluated on two public datasets (BCI Competition IV Dataset IIa and BCI Competition IV Dataset IIb), and the results are compared with the classification method of the state-of-the-art methods. Experimental results show that the proposed TSMFBEL algorithm can effectively construct a diversified ensemble classifier, and the average classification accuracy on the two datasets is 88.80% and 86.53% respectively, which are the highest among the state-of-the-art methods, and the standard deviation of the results is also the lowest. Excellent classification performance shows that the proposed algorithm has great potential in the decoding of MI EEG signals.