Research On Motor Imagery EEG Pattern Recognition Algorithms

Motor imagery brain computer interface is a very important brain computer interaction strategy.Its characteristic is that users control the robot or external machine by motor imagery related brain signals.Because of its great potential application in rehabilitation of motor function,it has attracted wide attention.Brain computer interface system provides a novel way of communication between the human brain and the external world.It is a collection of hardware devices and software in itself.The important research content and key technology of the brain computer interface system is to identify the user's action intention from the detected brain neural electrical activity data and convert it into an execution command to control the external mechanical and electrical equipment through the appropriate pattern recognition algorithm.Therefore,it is of great importance to study the feature extraction,feature selection and feature classification of motor imagery EEG.Motor imagery brain computer interface system mainly uses amplitude modulation information of sensorimotor rhythm signal to reflect subjects' action intention.Such modulation usually produces event related desynchronization or event related synchronization phenomenon.For different subjects,event related desynchronization and event related synchronization often occur in different brain regions,different frequency bands,and different time periods.In this paper,a frequency domain optimization and global parameter optimization algorithm in time domain is proposed based on dual tree complex wavelet transform and particle swarm optimization algorithm.However,in this algorithm,optimizations in frequency domain and time domain are relatively independent.On this basis,this paper further proposes a joint optimal parameter selection algorithm based on artificial bee colony algorithm in time and frequency domain.The common spatial pattern algorithm has been proved to be one of the most effective algorithms for motor imaginary EEG feature extraction,but the channel selection,frequency band filtering and time interval selection have a decisive influence on the effectiveness of the algorithm.In addition,when calculating the covariance matrix,it is easily disturbed by noise samples or abnormal samples.Furthermore,when designing the optimal spatial filter,most of the researches select the filter vectors directly based on the eigenvalues.To solve the above problems,this paper proposes the idea of sparse regression to optimize multiple local time frequency blocks.We also analyzed and studied the robust spatial filter bank design to reduce the influence of EEG non-stationary characteristics and noise on the common spatial pattern algorithm.In the sparse representation classification algorithm,the test sample is linearly represented by all training samples,which increases the flexibility of the adaptive operation of the algorithm and raises the higher requirements for the quality of the training samples.For many complex reasons,such as the non-stationary characteristics of EEG or the misoperation of subjects,some abnormal samples may be mixed into the dictionary.In order to remove the abnormal samples in the dictionary and improve the classification performance of the sparse representation classification algorithm,a dictionary cleaning scheme based on the k nearest neighbor algorithm is proposed.We collected the right hand index finger motor imagery EEG dataset to verify the algorithm,and the experimental results confirmed the identifiability of the motor imagery EEG of the index finger and the effectiveness of the improved algorithm.In the stage of motor imagery EEG feature classification,linear discriminant analysis algorithm is frequently applied to brain computer interface system.But from the related literatures of brain computer interface system,kernel methods often get better classification results than linear methods.In this paper,the idea of multi kernel discriminant analysis is introduced into the classification of motor imagery EEG signals.A multi-kernel construction scheme based on linear kernel and Gaussian kernel is proposed,and the Fisher 's linear discriminant analysis criterion is used as the basis for weight selection of kernel function.For feature optimization,a multiple local time frequency block optimization scheme based on composite kernel support vector machine algorithm is proposed in this paper.In the traditional motor imagery EEG pattern recognition algorithm,the two stages of feature extraction and feature classification are relatively independent.Although this traditional pattern recognition method has been widely used,the manual selection of EEG features depends on experience and prior knowledge,and feature extraction algorithms and feature classification algorithms often use different target functions.These will affect the effect of pattern recognition in a certain degree.In this paper,a deep learning and pattern classification algorithm for motor imagery EEG of limb movement is proposed.A multi-layer convolution neural network model is designed for motion imaginary EEG pattern recognition,which directly extracts the deeper and more interpretable features of the original EEG,and the feature extraction and feature classification are included in a framework to avoid the loss of information caused by the separation of the two stages.The research of this paper enriches the idea of motor imagery EEG pattern recognition,and can solve some of the key problems that exist at present,and can also be used for reference to the problem of pattern recognition in other fields.