Recognition Of BCI Motion Imagination By Co-training Based Domain Adaption Algorithm

Brain Computer Interface(BCI)technology,as a hot topic in the field of artificial intelligence,has been applied to multiple fields.Its development needs to rely on the research and analysis of EEG signals However,EEG signals themselves have the characteristics of noise and instability,and in real life,EEG signals are mostly unlabeled data,making it difficult to classify them Based on this,this paper uses the domain adaptation method in transfer learning to classify EEG signals.This method can solve the problem of insufficient tag data when the data distribution of source domain and target domain is inconsistent but the classification task is consistent.In addition,incorporating collaborative training on this basis can achieve better classification results.The main work of this thesis is summarized as follows:Firstly,the EEG signal is preprocessed.Specifically,the data set is aligned in Euclidean space(EA)to make the data distribution in different domains more consistent.regularization Common Spatial Pattern(RCSP)is used to extract features from EEG data.Secondly,due to the fact that the domain adaptation process for discriminating joint probabilities is based on pseudo labels in the target domain,and simple classifiers have lower accuracy in classifying unlabeled data,this paper considers adding pseudo labels to unlabeled data through the process of adding co training to improve the classification accuracy of the domain adaptation process And experiments were conducted on two datasets of motion imagination,data sets 1 and data sets 2a,to verify that the domain adaptation process incorporating collaborative training has better classification performance than the process directly applying domain adaptation,with classification accuracy improved by 8.72% and 3.9%,respectively.Finally,a deep domain adaptation model based on task decomposition is proposed based on the idea of incorporating collaborative training for domain adaptation by decomposing the domain adaptation task into two subtasks: semisupervised learning(SSL)in the target domain and unsupervised domain adaptation(UDA)across domains,two models are trained on two subtasks,and the two models cooperate with each other in each iteration Transforming the problem of mutual collaboration between traditional collaborative training views into mutual collaboration between two task models on a single view can compensate for their shortcomings and achieve better prediction results compared to simply adding traditional collaborative training processes Experiments were conducted on datasets data sets 1 and 2a to verify that the domain adaptation process based on task decomposition deep collaboration has better classification performance than the domain adaptation process solely incorporating collaborative training.The classification accuracy has been improved by 1.96% and 1.65%,respectively.Compared with the classification methods proposed in recent literature,the classification accuracy has been improved.