Research On Classification Method Of Motor Imagery Based On Modified S Transform And Enhanced Convolutional Neural Network

Brain-computer interface(BCI)technology uses professional equipment to collect EEG signals,and then performs signal preprocessing,extract signals feature and pattern recognition to control external equipment.From the 1970 s to the present,with the in-depth understanding of the nervous system and the rapid development of computer technology,BCI has played an irreplaceable and important role in the fields of biomedical engineering,computer science,communication and even artificial intelligence.As the signal carrier of BCI technology,EEG signals is a kind of physiological signal with spatio-temporal characteristics,among which the classification and identification of motor imagery EEG signals is one of the core branches of BCI.The brain-computer interface paradigm based on motor imagery mainly relies on the physiological basis of event-related synchronization and event-related desynchronization to distinguish different brain inten-tions,and realize the purpose of "human-computer interaction" by controlling external devices.It can be seen that the effective feature extraction and accurate classification of motor imagery EEG signals are very important,which directly affects whether the brain’s intention can be converted into correct commands and achieve effective control of external equipment.This paper is based on the traditional EEG signals classification and recognition research,and the research work is summarized as follows:This paper proposes an EEG signals classification model based on the combination of modified S transform(MST)and enhanced convolutional neural network(ECNN).In order to select the optimal feature extraction algorithm,this paper compares the effects of shorttime Fourier transform,wavelet transform,S transform and modified S transform on the time-frequency feature extraction of motor imagery EEG signals.The Re LU activation function is greatly affected by the learning rate and is prone to over-regularization.An enhanced convolutional neural network is proposed,which uses the SELU activation function with self-normalization properties and uses different regularization parameters for different layers to obtain better training effect.Further,aiming at the problems of insufficient EEG data and redundant multi-channel EEG information,a data processing method of data amplification and channel selection is proposed,and the MST-ECNN model based on data amplification and channel selectionbased model are built.The MST-ECNN model.The sample size of the original data is increased by data augmentation,and redundant irrelevant channels are removed by the wrapping channel selection method to improve the classification efficiency and accuracy of the model.Finally,the effectiveness and superiority of the above methods are verified using 3different datasets.On the basis of classification accuracy,6 evaluation indicators including sensitivity,specificity,positive predictive value,area under the curve and ROC curve were introduced to further optimize the evaluation system.