Research On Classification Of Motor Imagery EEG Signals Based On Deep Learning And Attention Mechanism

Feature extraction and classification of Electroencephalogram(EEG)signals is an important part of the Brain-computer Interface(BCI)system.Motor-imagery(MI)BCI system can convert EEG signals into control instructions for auxiliary equipment,providing a rehabilitation method for motion-disabled patients,which has high practical value.However,how to accurately extract temporal-spatial features of EEG signals to improve the classification accuracy of MI BCI is still a problem that needs to be solved urgently.To this end,this study uses deep-learning techniques to classify two types of hand(Left-hand and right-hand motion)MI EEG signals in the GIGA public-available dataset.The main work of this paper is summarized as follows:(1)First,based on the frequency characteristic of MI EEG,the band-pass filter of 8Hz-30 Hz was used to filter EEG signals.Then,common average reference and independent component analysis were used to remove noise and artifact.Last,EEG signals were timesliced and down-sampled.Two types of hand MI EEG datasets were obtained,which laid the foundation for the subsequent classification research.(2)Classification of MI EEG signals based on deep-learning methods.EEGNet and EEG-Inception models were used to classify two types of hand MI EEG signals.To address the inadequate classification performance of EEGNet and EEG-Inception models,a classification model based on convolutional neural network(CNN-Net)was designed.Lefthand and right-hand MI EEG signals were classified using four pairs motor-related EEG channels.The experimental results show that CNN-Net outperforms EEGNet and EEGInception,and achieves better classification effect.And by comparing existing research methods,the effectiveness of the CNN-Net model was further verified.CNN-Net has the best classification effect when using 4-channel(FC3-FC4,C1-C2)signals,and the average accuracy rate of classification reaches 94.90%.(3)To explore the effect of attention mechanism on performance of CNN-Net model,SENet(Squeeze-and-Excitation Network),ECA(Efficient Channel Attention),and CBAM(Convolution Block Attention Module)attention mechanisms were respectively incorporated into CNN-Net model designed in(2).Influence of different attention mechanisms on the classification performance of CNN-Net were comparatively analyzed.The experimental results show that the integration of attention mechanism into the middle layers of the network achieves better classification results than the shallow and deep layers of CNN-Net.Under the optimal incorporation method,the CNN-Net model incorporated with SENet achieved an average accuracy of 95.79% in classification.Compared with the CNN-Net model,its classification performance was further improved.Its classification performance is better than most existing methods and avoids the complex feature extraction process.These results provide new references for the classification of motor imagery EEG signals.