Study Of Domain Adaptation Algorithms For MI-EEG Signal Based On Covariance Matrix

The statistical characteristics of training samples and test samples is different in BCI experience of the same user because of the changes of EEG signal,which decrease the performance of classifiers like Support Vector Machine in BCI system.A common solution is to perform long-term offline training before each use,which greatly limits the availability of MI-BCI system.To solve the problem,this thesis introduces the domain adaptation method in transfer learning into the MI-BCI system,and proposes a domain adaptation algorithm based on the covariance matrix difference to make the distribution characteristics of the different data mentioned above more consistent.The metric of covariance matrix difference is specifically designed under the following three specific situation:Firstly,in the case of a single user's single motor imagery session,this thesis designs KLDA-WT algorithm,a revised version of the KL divergence based KLDA algorithm.KLDA-WT uses the output probability of the SVM to weight the test samples,and introduces the threshold based on the 2-norm as the criterion,so that it can be used for sample imbalance and filter the test samples with large changes.Compared with the algorithm without domain adaptation,all users in this thesis can maintain the accuracy using the KLDA-WT algorithm,and achieve an accuracy improvement of up to 8% for individual users.Secondly,in the case of a single user's multiple motor imagery sessions,we research on the S2norm-Wtr algorithm,which employs the covariance matrix 2-norm as the metric of covariance matrix difference instead of KL-divergence.The new algorithm solves the difficulty caused by the non-conductible of KL-divergence,and admits the further easy improvement of the algorithm.The algorithm constrains the trace of the mapping matrix and reduces it to the local optimum problem with the box constrained elements.Then the objective function is weighted by the stationarity and importance of each element in the covariance matrix.The simulation results show that compared with the algorithm without domain adaptation,S2norm-Wtr can improve the accuracy of users with an average increase of 5.87% and a maximum of 15.13%,except A06.In the small sample case,the result of S2norm-Wtr is the same with that of re-offline training,indicating that it is feasible to replace the offline training process with domain adaptation method.Finally,for the problem that the statistical characteristics of key elements in the covariance matrix of individual users are unstable,this thesis designs a covariance difference metric based on exponent,which uses a small range to replace the average of the training sample covariance to calculate the mapping matrix.The results show that the EXP-Wtr algorithm can improve the classification accuracy of the above userss.In the covariance-based domain adaptation algorithm proposed in this thesis,we design different difference metrics of covariance matrix for the above three types of problems,which makes the domain adaptation algorithm more stable,the overall classification accuracy better,the offline training time less and the algorithm more applicable in the BCI system.