The Classification of Motor Imagery EEG Signals Based on the Time-Frequency-Spatial Feature

The effective features of the motor imagery (MI) electroencephalogram (EEG) signals plays a significant role to improve the classification accuracy for the brain-computer interface (BCI) system. Some traditional methods usually extract the frequency or spatial features without considering the related information between different channels that would affect the classification performance. This paper proposes a new method for feature extraction of EEG signals based on the fusion of time-frequency and spatial features. At the beginning, the common spatial pattern (CSP) algorithm is adopted to extract the spatial features. Then the discrete wavelet transform (DWT) and the wavelet packet decomposition (WPD) are used to extract the µ rhythm of the motor imagery EEG signals as the time-frequency features. After that, by combining the spatial and time-frequency features, the time-frequency-spatial feature is formed. Based on different kinds of features, the experimental data are classified by using the support vector machine (SVM), as well as the sparse representation classification (SRC) algorithm with the elastomeric network (EN) and L1 norm, respectively. The experimental results show that the SRC with EN has a better performance on either the time-frequency feature or spatial feature than the SRC with L1 norm does. In contrast, the SVM and the SRC with Ll norm perform better than the SRC with EN based on the time-frequency-spatial feature. The study concludes that the time-frequency-spatial feature cooperating with the certain classifiers can achieve the good classification effect for the MI EEG signals, which not only reduces the operation time but also improves the classification accuracy.