A Novel Motor Imagery EEG Classification Model Using Frequency-Temporal-Spatial Convolutional Neural Network with Channel Attention

Abstract

Motor imagery (MI) EEG classification is a crucial component for brain-computer interface (BCI) system, the quality of features extracted from EEG signals greatly affects the MI classification performance. Nevertheless, most researches caused information loss while performing EEG feature extraction. To extract multilevel recognizable EEG features and improve classification performance of MI tasks, this paper proposes a novel MI EEG classification model of frequency-temporal-spatial multi-layer CNN with channel attention. In this model, multiple convolution kernels with different size are used to extract features at relevant frequency bands, the squeeze-and-excitation block based channel attention mechanism is added to learn distinct frequency characteristics, then via layer-by-layer convolutional operation, EEG features from frequency-temporal-spatial can be excavated. The model can learn EEG features of three domain and classify MI tasks simultaneously. Finally, the proposed method is evaluated on two BCI datasets. For dataset BCI-C IV-2a, the average classification accuracy and kappa value are 89.14%, 0.85, respectively; for dataset BCI-C IV-2b, the average classification accuracy and kappa value are 89.19%, 0.78, respectively. Additionally, when comparing to other MI classification methods on the same EEG datasets, the classification performance is superior to state-of-the-art methods, the effectiveness and superiority of the proposed method is validated.