Application Of Multi-classification Algorithm In MI-EEG Signal Recognition

Brain-computer interface(BCI)is a technique that allows EEG signals to control and communicate with external devices without relying on their own nerve and muscle tissue.In a paper published in Nature in 2019,people actually turned “mind-reading” into reality by translating the sounds of electrical brain signals.In the future,BCI has a broad application prospect in the field of intelligent home life,military,education and rehabilitation medical fields.Based on the research and understanding of the brain,the study of EEG signal is also gradually applied to the diagnosis of brain tonic diseases and the interventional treatment of neurological diseases.Based on the motor imagery of BCI system takes the EEG signals generated by the subject's motor imagination as input,and realizes the communication and control with the external devices through computer processing,so that their intentions can be understood by people.The signal processing part of the BCI system is mainly to process EEG signals into commands that the computer can recognize and execute.How to quickly extract and identify the information carried in EEG signals is an important topic in BCI system research.Based on the 2008 BCI competition IV of dataset 2a,this paper designed a signal processing system based on four class EEG signals.(1)Pre-processing used Independent Component Analysis(ICA)to filter out the doped artifact signals(EOG,ECG and EMG)and other noise interference,and enhanced the frequency domain characteristics of EEG signals.Stockwell transformation is used to transform the signal into time-frequency,analyze the time-frequency domain energy changes of EEG signals,capture useful EEG information,and analyze the frequency domain characteristics of EEG signals.in frequency domain.According to the time-frequency information of EEG signals in ST,the power spectral density(PSD)of each channel based on frequency domain was obtained,and the PSD of all channels jointly constituted the characteristic graph of Convolutional Neural Networ(CNN).(2)The EEG signals collected in the data set are continuous,which include a large number of useless EEG signals in addition to the signals containing motion imagination experiments.The complex experimental process makes the useful samples of EEG signals less.Convolutional neural networks usually deal with large sample problems.Small samples make it difficult for the neural network to learn useful features of the signal,making it difficult for the signal processing system to converge.In order to solve the problem of insufficient experimental data,this paper uses two methods of adding sliding time window to signal and adding Gaussian white noise to realize data enhancement,which provides sufficient data support for the later deep learning algorithm.(3)The feature map constructed by S transform contains the time-frequency domain and spatial features of EEG signals.The feature extraction part uses the convolutional and pooling layer of CNN to extract time-frequency and position information of EEG signals.Based on the CNN designed in this paper,we proposed a parallel CNN framework.Multiple CNN are trained side by side to extract the features of different feature maps and all feature information is fused using a fully connected layer.Finally,support vector machines(SVM)are used to implement feature classification.The classification accuracy of PCNN and CNN is 83.2% and 84.1%.The classification accuracy of the EEG signals of the PCNN structure is better than that of CNN.The training time of PCNN is 5 times faster than that of CNN.The classification accuracy of PCNN-SVM is 86.6%,which is higher than the classification result of PCNN.(4)The redundant features carried in the EEG signals affect the classification performance of the system.In this paper,the Bat algorithm is used to optimize the feature weights of the PCNN fully connected layer fusion,and the optimal feature subset is selected to improve the classification performance of the system.The signal processing framework proposed in this study improves the classification performance of four types of EEG signals,which provides a theoretical basis for the classification of multiple types of EEG signals,and also provides ideas for the research and application of BCI systems based on motor imagery.