Research on Feature Fusion Based on CNN-LSTM Network for Motion Imagery EEG Classification

By analyzing the time-domain and spatial domain features of EEG signals, a feature fusion based on CNN-LSTM parallel network for EEG classification is proposed, which realizes feature fusion for different networks and different layers. The parallel network consists of two layers of CNN and one layer of LSTM to decode EEG signals. The CNN network is used to extract the spatial features of signals from different channels, and the LSTM network is used to extract the temporal features of signals, so as to extract the features of original EEG signals from the spatial and temporal domains. At the same time, the features of the intermediate layer of the flatten layer storage network are also used. The classification results can be obtained by fusing at the full connection layer. Public Dataset BCI competition IV-2a are used to verify the proposed algorithm. The Kappa value and average correct rate obtained in Datasets 2a data set reach 0.824 and 87.6% respectively, which are 9.5% and 3.7% higher than those of EEGNet algorithm and higher than the existing base-line classification algorithms.