Research On Small Sample MI-EEG Signal Processing Algorithm Based On Domain Adaptation

The brain computer interface of motor imagery system has the advantages of spontaneity,continuity,and flexible application methods.However,the poor signal reusability caused by the non-stationary characteristics of the motor imagery EEG signal and the experimental operation process leads to a long-time offline training phase before each online MI-BCI system experiment,which greatly affects the practicability of the online MI-BCI system.In the traditional MI-EEG signal processing,the insufficient number of training samples will lead to overfitting of classifier model,which has a negative impact on the classification accuracy.In order to solve the contradiction between the offline training time and the number of training samples,this paper combines the domain adaptation ideas in transfer learning into the processing of MI-EEG signals.Through the participation of auxiliary subjects,the algorithms are designed respectively from the two perspectives of selectively increasing the number of training samples and optimizing the feature distribution of the small samples themselves,which improves of classification accuracy under the condition of small sample training.Firstly,based on the combination of band-pass filter,common spatial pattern and support vector machine algorithm,combined with open source data,this paper simulates and analyzes the effect of the change of the number of training samples on the classification accuracy of testing sample.By changing the number of training samples,it is concluded that the classification accuracy will increase with the increase of the number of training samples,which confirms the view that the classification accuracy is not good under the small sample training condition.Meanwhile,it is found that with the increase of the number of training samples,the classification accuracy of different subjects increases by different degrees,and there is a certain upper limit.Secondly,under the condition that the domain adaptation needs to be met,from the perspective of selectively increasing the number of training samples,this paper designs a domain adaptation algorithm based on sample selective mixing and classification prediction label weighting.In this algorithm,the samples of single auxiliary subject are directly mixed with the training set samples of target subject.After feature extraction and self-classification,the misclassified samples of auxiliary subject are deleted,and then the testing set samples of target subject are classified and predicted.Finally,based on the classification prediction label,two weighting methods are designed,which are based on the confidence probability of classification prediction label and the contribution degree improvement based on the classification accuracy rate.In the simulation results,the algorithm can improve the classification accuracy under the condition of small sample training by 7.62% ?21.94%.Finally,from the perspective of optimizing the distribution of the small samples own features,based on the Kullback-Leibler divergence weighted composite common spatial pattern algorithm proposed by others,aiming at the defects in the calculation process of KL divergence,the calculation method of the auxiliary weight is changed,and the domain adaptation algorithm based on the direct contribution of classification accuracy improvement weighted CCSP and the indirect contribution of classification accuracy improvement weighted CCSP are proposed.The two algorithms take the difference of the classification accuracy between the samples of auxiliary and the samples of non-auxiliary as the contribution of classification accuracy improvement,and calculate the weight value.The difference is that the former algorithm is to mix the samples of a single auxiliary subject with the training set samples of target subject to directly calculate the difference of classification accuracy.While the latter algorithm combines the influence of the number of auxiliary subject samples on the classification accuracy.By deleting the samples of each auxiliary subject in the auxiliary sample set in turn,the difference of classification accuracy after the remaining samples are mixed with the target subject training set samples at the same time is regarded as the contribution of the indirect classification accuracy.In the simulation results,the former algorithm can improve the classification accuracy rate under small sample training conditions by8.40%?21.63%,and the latter algorithm can improve 9.59%?24.09%.For different subjects,the specific promotion values are different,and the overall performance of the two algorithms are better than that of the KL divergence weighted algorithm.In this paper,according to the measures to optimize the performance of different classifier models and reduce the occurrence of over-fitting,the three algorithms proposed can greatly improve the classification accuracy under the condition of small sample training,that is to achieve the purpose of reducing the time of the offline training phase and enhancing the practicability of the online MI-BCI system.