Research On Error-related EEG Signal Classification Algorithm Based On Riemannian Geometry And Joint Improvement

The brain-computer interfaces(BCI)is independent of peripheral nerves and human muscle,it provides a new way for the users to communicate with the external environment.BCI devices can not only help people with disabilities return to normal life,but also promote the development of military,production,education,household,transportation and so on.Error-related signal(ErrP)is the signal that describes the changes in EEG information after a user perceives the wrong thing.ErrP is the key to identify misunderstood commands and is of great significance for medical equipment.It directly related to the flexibility and safety of the device.However,due to the poor quality,instability and the obviously individual differences of the currently EEG signals,the model trained by original signals has poor generalization performance on new subjects.Moreover,there are lots of electrode channels and a fewer data that lead to overfitting.In order to overcome these difficulties,we take advantage of Riemannian manifold learning and transfer learning methods: Riemannian geometry is able to fit data high-dimensional features;transfer learning can reduce the dependence on the target data and computation.The main research work of this paper are as follows:An ErrP signals recognition algorithm based on Xdawn and Riemannian manifold learning is proposed in this paper.First,we process the signal by means of Xdawn spatial filter.Then utilizing the Riemann kernel function to calculate the symmetric matrix of the sample covariance matrix in the tangent space at the Riemann mean point.Finally,half-vectorizing the matrix to extract the Riemannian features.After adding meta features and leak features,we use bagging algorithm and elastic network to detect ErrP.Compared with other methods,a series of experiments on public datasets will prove the effectiveness of this method.In this paper,a joint improvement algorithm is proposed for ErrP singals recognition.First,replacing the Xdawn filter with the Riemann distance spatial filter which has add the average evoked potentials,this way can preserve the correlation of the signal while reducing the amount of data.Second,we analyze the frequency bands of the ErrP signals so that we can design the filter bank’s structure and extract the features of multi-frequency bands.Finally,the features are aligned.This way makes the features distribution of the source domain and the target domain more consistent.The improvements in the above three aspects are integrated into a joint algorithm,and then experiments are carried out.Through the analysis of the results,we can find the joint algorithm proposed in this paper brings about the improvement of performance;meanwhile.In addition,it provides a new explanation for the brain functional area that induces the ErrP signals.