The Research On Transfer Learning Optimization Algorithm For Motor Imagery-Based Brain-Computer Interfaces

The Brain-Computer Interface(BCI)can establish a system for simulating peripheral nerve-controlled muscle tissue,which decodes electroencephalography(EEG)signals collected from the human scalp into control signals.Thus BCI may provide a pathway between patients with motor disabilities and the environment,and control the external equipment to realize the function of the limb.This new type of interactive system is at the forefront of priority development and research in the field of neural engineering,which has important scientific significance and application prospects.BCI had been applied in therapeutics for decades.However,because the training model of BCI system requires a large amount of sample data,the subjects will feel tired after a long time of training,which will affect the quality of signals,thus restricting the advance and application of BCI,especially for the BCI based on motor imagery(MI).This paper proposes the transfer learning algorithms to solve the problems of insufficient training data and the non-linear and non-stationary EEG signals.First,we proposed a weighted covariance matrix method based on Common Spatial Pattern(CSP)to supplement the training trials of the target subject.The similarity between the existing source subject data and the target training data is used as a weighted item.The target subject's CSP feature matrix is-weighted and combined to update the feature data,and the target subject data is combined with the supplementary training data,to achieve the data transfer of other subj ects with the same task to reduce training samples.Then,for the problem of non-stationary and non-stationary EEG signals,a feature transfer method based on domain distribution adaptation is proposed and applied to the classification of motor imaging data.Because the subjects are easily affected by their own mental state and environment,the distribution of EEG signal data changes and the difference between different subjects is large.It is impossible to directly add source domain data for training.The transformation matrix is used to map the feature data into a new feature space,so that the edge distribution and conditional distribution of the source and target data after the mapping are close to achieve the purpose of feature sharing.The results show that the proposed method is helpful to improve the classification accuracy of the subjects' motor imagination to a certain extent.?By using the experimental data of other subjects,the training data required by the target subjects to train the classification model is reduced,and the classification accuracy is improved,indicating that the transfer learning algorithm has better adaptability characteristics than the traditional machine learning method.