Research Of Classification Algorithm Of Brain Signal Based On Transfer Learning And Discriminative Feature Learning

Brain-computer interfaces provide a new communication way for user and the surrounding environment without using pathway of nerve and muscle.With the rapid development of electronic technology and artificial intelligence,brain-computer interfaces successfully applied on lots of fields,including medicine,military,smart home and game.However,due to the nonstationarity of the signal and large variation between brain signal of different subjects,performance drop occurs when directly used the model trained on old data to predict new data.Consequently,long-time calibration is a necessary process before using the system,which has become a major bottleneck in the development of brain-computer interfaces.To tackle this obstacle,one promising approach is transfer learning,which use data from other subjects to construct an auxiliary dataset for target subject.The target data combined with auxiliary dataset could learn a better model.The main research work are as follows:A class centroid matching transfer learning algorithm based on feature extracted from Riemannian tangent space is proposed.First of all,the covariance matrices are computed as raw features.After that,feature vectors can be acquired by projecting feature matrices into Riemannian tangent space and half-vectorizing the projected matrices.To make full use of structural information in target domain,the class centroid matching transfer learning algorithm can be applied on vector-based feature space.In this algorithm,due to the task is a binary classification problem,a pair of class centroid can be inferred with the help of clustering.Optimal subspace can be obtained by minimizing the distances between source class centroid and target class centroid.Furthermore,a discriminative loss added in total loss function can enhance the feature separation between classes.A series of experiments on an magnetoencephalogram dataset show that the proposed algorithm actually improves the classification accuracy.Both traditional feature extraction and deep feature extraction are important in the field of brain information processing.It is researched that the distribution of features learned in deep learning varies from source domain to target domain.This dissertation proposes a deep domain adaptation framework in which the domain confusion can be maximized and feature discriminability can be maintained.Two strategies of discriminative feature learning are introduced to the deep domain adaptation network.One is based on instance and the other is based on class center.Several cross-subject electroencephalogram classification experiments show that the discriminative feature learning improves the performance of error-related negative potential classification.