Research On Motor Imagery EEG Signal Recognition Method Of Brain Controlled Assistive Device

Brain computer interface(BCI)based on motor imagery(MI)has attracted extensive attention of a large number of researchers around the world because it can directly establish the communication between the brain and external electronic devices without the participation of peripheral nerves and limbs.Brain computer interface can convert motion intention into the corresponding control command to control external device by analyzing the inherent of the electroencephalogram(EEG).Most of the existing studies focus on the decoding of different limbs,such as left hand,right hand,foot and tongue,and have achieved good recognition results.However,with deeper research of the problem,a few scholars have carried out the decoding of different MI tasks of the same joint to enable users to intuitively control assistive devices.However,those studies have the low classification accuracy.Based on the analysis of lots of studies,from the perspective of traditional machine learning,the decoding method of multi-class MI tasks of the same joint is carried out.In the process of deep research,considering the high requirements of traditional machine learning for Feature Engineering,the corresponding MI decoding method based on the deep learning theory is explored.A brain computer interface system is established to naturally control the biomimetic manipulator,which proves the practical significance of this paper.The main contents and innovations are as follows:(1)A multi class MI tasks recognition method based on LMD-CSP and MOGWO-TWSVM is proposed.Firstly,the preprocessed EEG signal is decomposed into a series of product functions(PFs)by local mean decomposition(LMD).Secondly,the effective PF components are selected according to the entropy(En)and supper entropy(He)of cloud model.Then,the selected PF components of each channel are reconstructed into a new signal matrix,and common spatial pattern(CSP)is used to extract MI features.Finally,the parameters of twin support vector machine(TWSVM)is optimized by multi-objective grey wolf optimizer(MOGWO)to classify MI features.To test the effectiveness of the proposed method,the proposed method is compared with the existing feature extraction,classification and overall recognition method.(2)A multi class MI tasks recognition method based on symplectic geometry mode decomposition(SGMD)and data augmentation-based long short-term memory(DA-LSTM)is proposed.Firstly,the SGMD is applied to remove the noise of the preprocessed EEG signal,where the energy proportion of ? rhythm and ? rhythm is introduced as the termination condition.Secondly,the time and frequency features of the denosied EEG signal are extracted and reconstructed into a two-dimensional feature matrix.Finally,the feature matrix is input into the deep convolution generative adversarial networks(DCGAN)to expand the training set,and then the long short-term memory(LSTM)is further trained to recognition MI tasks.Experimental results show that the proposed method can effectively identify shoulder abduction,flexion and extension MI tasks,and has more development potential.(3)A BCI system of brain-controlled assistive devices is built.Firstly,the system framework is deigned.Secondly,the upper computer software is developed using MATLAB and LABVIEW,and STC89C52 is applied as the control core to design the corresponding hardware.Finally,the MI EEG signal of shoulder flexion and extension is used to intuitively control the biomimetic manipulator.Through the offline measurement experiment,the effectiveness of the proposed method is verified,and the practical significance of this paper is also proved.