Research Of Motor Imagery Recognition Based On Attention Mechanism And Dense Neural Network

The challenge for motor imagery(MI)in brain computer interface(BCI)systems is finding a reliable classification model that has high classification accuracy and excellent robustness.The main limitations to the performance of motor imagery EEG are low signalto-noise ratio and the non-smoothness of electroencephalogram(EEG).To address this dynamic complexity of the EEG that makes decoding motor imagery EEG signals difficult,this paper focuses on a study of signal recognition based on attentional mechanisms and dense neural networks.The aim of this paper is to explore a general framework based on time-series classification that can advance the practical process of MI-BCI.First of all,to address the low signal-to-noise ratio of EEG signals,band-pass filtering,common mean reference and independent component analysis are used to perform preprocessing operations to ensure the quality of motor imagery EEG signals.Furthermore,considering the channel correlation of EEG,a novel three-dimensional representation of EEG signals is designed to solve the problem of traditional two-dimensional arrays not being able to capture the feature information of adjacent electrodes,laying a solid foundation for accurate recognition of EEG signals.Secondly,considering that most current methods have the problem of difficulty in taking into account the multidimensional feature information of EEG signals,a parallel spatial-spectral-temporal attention feature extraction method is proposed for dynamically capturing the signal features of different brain regions,frequency bands and time.The extracted feature maps are also classified by means of shallow neural networks,and the method is shown to be able to extract distinguishing features by comparison with existing advanced signal recognition methods.Morever,a densely connected neural network-based classification and recognition method is proposed to address the problem of low accuracy of EEG signal recognition due to large individual differences in EEG signals.The advantage of the dense connectivity of the network and the cross-stage local structures is taken to integrate the spatial-spectraltemporal feature information of EEG signals into a unified network architecture,which improves the computational efficiency of the model and its classification performance.The cross-subject classification training strategy is designed to further validate the effectiveness and generalisation ability of the model,in view of the strong individual dependence of EEG signals.Finally,by combining the above studies,the signal recognition models constructed in this paper have obtained better results than most current signal recognition methods on the BCI IV 2a dataset and high-gamma-dataset,with average accuracies of up to 84.45% and97.88%.In addition,based on the model architecture trained from offline experiments,MIEEG online recognition experiments are designed to achieve accurate recognition of four types of motor imagery states for different subjects.The results obtained validate the feasibility of the theoretical approach of this paper and provide a practical basis for the practical research of the MI-BCI system.