| The rapid development of artificial intelligence promotes the progress of brain-computer interface(BCI)technology.The research of motor imagery(MI)based BCI can help individuals with motor disabilities contact with the external environment.However,there are still some problems in decoding motor imagery electroencephalogram(EEG)signals,such as low decoding accuracy and poor adaptability.Therefore,this study proposes a MI-EEG classification algorithm based on time-frequency common spatial pattern(TFCSP)and attention-network to address the above problems.The main work of this paper includes:(1)In the field of motor imagery EEG decoding,common spatial pattern(CSP)is a common and effective feature extraction method,but CSP feature extraction method only considers the spatial information of EEG signals,ignoring the time-frequency information.Therefore,this study proposes TFCSP feature extraction method,which divides time and frequency bands of EEG signals before extracting CSP features,and then extracts CSP features for each time-frequency band.Compared with the traditional CSP feature extraction method,TFCSP method makes full use of time and frequency information of EEG signals and improves the decoding accuracy.(2)The traditional convolutional neural network has poor interpretability in decoding EEG signals.To address the above problem,this study combines attention mechanism with convolutional neural network to enhance the interpretability of the model.In addition,the visualization maps of attention mechanism show that attention mechanism can identify the active EEG channels of different subjects’ when imaging moving,and help neural network to classify by enhancing the weight of their active EEG channels.(3)In order to improve the model’s utilization of the time,frequency and space information of EEG signals and enhance the robustness of the model,this study proposed the MI-EEG classification model combining the TFCSP features with the attention based convolutional neural network,which further improved the model’s performance.We evaluated the algorithm on two public MI-EEG datasets,BCI Ⅲ Ⅲa and BCI Ⅳ Ⅱa,achieving the average accuracy 86.39%and 79.28%respectively,which proved the robustness and effectiveness of the model,and also proved that the model has a good ability to extract time,frequency and space features. |
