Study On Motor Imagery Based Brain Computer Interface Based On Extreme Learning Machine

Brain-Computer Interface(BCI)as a new human-computer interaction system,has great practical value and broad application prospects in the fields of rehabilitation training,game entertainment,education and military.This technology has gradually become a research hotspot in the world.The research subject of this thesis is BCI system based on motor imagery.Feature extraction and classification recognition of motor imagery Electroencephalogram(EEG)signals is an important step to realize this type of BCI system,meanwhile designing an effective classifier with high classification accuracy and strong generalization capability is essential to achieve high performance.This thesis studied the ELM based classification method of motor imagery,since the Extreme Learning Machine(ELM)has the advantages of strong generalization ability and fast learning speed.Firstly,a supervised classification model based on ELM was proposed,which derives from the ideas of the optimization and ensemble learning.Then,a semi-supervised classification model based on ELM for online application is proposed in the case of small training samples,such that a motor imagery BCI system is designed.The main research contents of this thesis are as follows:(1)A supervised classification method of parameter optimization extreme learning machine based on ensemble learning was proposed.The classification accuracy of ELM may not be high and stable,Due to the randomly generated input weights and hidden biases.In order to overcome the problem caused by randomly generated parameters,this classification model first uses particle swarm optimization to optimize the input weights and hidden biases of ELM simultaneously,then multiple optimized ELMs used as the base classifiers are fused into a strong classifier using majority voting strategy to improve the classification performance of motor imagery BCI.Two public available BCI competition datasets were used to verify the two-class and four-class experiments.Experimental results showed that the new classification model can significantly improve the classification accuracy of motor imaging EEG signals.(2)A semi-supervised classification method based on improved regularized weighted online sequential extreme learning machine was proposed,which can achieve online classification of motor imagery EEG signals with small training samples.The proposed method solves the problem of sample imbalance by weighting the online sequential extreme learning machine.Then,a linear classifier to assist the main classifier ELM in selecting unlabeled samples for updating the hidden layer output matrix is added to update the classification model.Two BCI competition datasets used as offline analyze to evaluate the proposed improved method.The experimental results showed that the proposed method can make a significant improvement in classification accuracy,and thus the effectiveness of the proposed method was also verified.(3)A motor imagery BCI system is devised and achieved.The classification model in the system respectively adopt the supervised and semi-supervised model studied in this thesis.The system realizes the collection and transmission of multi-channel EEG signals by mean of Neuroscan acquisition equipment.Then the upper computer software,written by Lab VIEW,was adopted to realize the functions of signal preprocessing,feature extraction,classification and recognition after receiving the signal.In order to evaluate classification performance of the designed motor imagery BCI system,three experiments were designed and implemented: 1)Subject training experiment.Collecting the EEG data through the designed motor imaging paradigm and training the initial classification model;2)PC software test experiment.Checking the functions of the online system PC software by the offline data;3)Pseudo-online experiment to evaluate system performance.Two subjects controlled the small lights by imagining left or right hand movements according to the selected path.The experimental results show that the BCI system designed in this thesis can achieve better classification results.