Motor Imagery EEG Intention Recognition Based On Interpretable Clustering Method

The electroencephalogram(EEG)signal obtained by non-invasive technology is widely used to control auxiliary equipment or rehabilitation equipment.So far,most BCI applications based on motion imagery EEG signals have used supervised algorithms to decode EEG information.However,the time-consuming process of training and calibration of the supervisory model limits the promotion of BCI applications.The clustering model is used to identify EEG intentions to reduce the training process of BCI applications.However,many clustering methods have encountered the development bottleneck of interpretability in their applications,and there is an urgent need to develop interpretable clustering analysis methods for BCI applications.Interpretable clustering can not only realize the classification of observation data,but also clearly explain the difference of each cluster.In this paper,we decode single trial motor imagery EEG signals based on interpretable clustering model.On this basis,a reconstruction feature space method is proposed to optimize multi-dimensional EEG features,and an optimization method is proposed to further improve the recognition performance of the interpretable clustering model.The main contents of the thesis are as follows:(1)A method based on the Discriminant Rectangle Mixed Model(DRMM)to decode single-trial motion imagery EEG signals is proposed.This method can reduce the training process of BCI applications,and at the same time overcome the disadvantages of the traditional clustering methods that give complex decision-making secondary boundaries,which cannot explain the differences between each cluster.It is proposed to use rectangular decision rules to define the boundaries of EEG features,and each rectangular decision boundary obeys Gaussian distribution.Since the decision rectangle is not differentiable,there is no way to optimize the rectangle decision boundary.This paper uses the sigmoid logic function to determine the boundary of the soft rectangle,and approximates the distribution of the rectangle decision rule through variational reasoning.Each rectangular decision rule can clearly explain the difference between the clusters,making the clustering results interpretable.First,the Common Spatial Pattern(CSP)is used to extract the multidimensional EEG feature from the public data set and auxiliary data sets.When the feature is reduced,the Fisher ratio is used as the cost function,and the projection direction corresponding to the maximum value of the Fisher ratio is automatically found as the best projection direction.After DRMM learns the optimal features,it obtains the rectangular decision rules for each class of EEG features and compares the DRMM interpretable clustering results with the standard models.Comparative experiments show that when it comes to recognizing EEG intentions,DRMM has comparable or even better recognition performance than standard clustering algorithms.(2)An optimal feature space reconstruction method is proposed to optimize the multidimensional EEG feature matrix to improve the separability of the optimal EEG features.Based on maximum separability,a discrimination criterion is established for multidimensional features and the maximum eigenvalue of the discrimination criterion is solved.At the same time,the eigenvector corresponding to the maximum eigenvalue is taken as the first best projection direction.The constraint condition that the eigenvector to be solved is orthogonal to the known projection direction is introduced,so that the solved eigenvector meets the requirements of space coordinate base.The multi-dimensional EEG features are projected to the optimal feature space composed by feature vectors to obtain more separable two-dimensional and three-dimensional EEG features.Finally,DRMM is used to identify the optimal EEG features.Comparative experiments prove that the feature optimization method proposed by this method has good performance,and the two-dimensional optimal feature has higher stability.(3)An optimized DRMM method is proposed to overcome the influence of anomalies on the interpretable clustering model.The unsupervised method is used to detects the outliers in the optimal features to overcome the shortcomings of Monte Carlo sampling that requires rich prior knowledge,and at the same time can avoid the time-consuming training process of the supervised model.First,a certain number of isolated trees are constructed,and based on the length of the search path of the feature in the isolated tree,an anomaly score formula is constructed to evaluate the anomaly degree of each sample.On the basis of determining the outliers,DRMM is used to obtain the rectangular decision rules of each class in the EEG features.The position relationship between the outliers and each decision rectangle is the main point,and the Euclidean distance is used as a supplement to complete the category prediction of the outliers.And on this basis,verify the optimal parameter selection problem in the optimization method.The experimental results prove that the optimization theory method is effective,and can further improve the accuracy of the DRMM decoded singletrial motor imagery EEG signal.