Research On Technologies Of Hybrid BCI Based On Transfer Learning

In recent years,with the rapid development of brain science and artificial intelligence,brain-computer interface technology has made breakthroughs in both theory and application,and is considered to be one of the most important cuttingedge technologies in the twenty-first century.The hybrid brain-computer interface has become a more powerful artificial intelligence system than the single-mode brain-computer interface system due to its more reliable,stable and multidimensional control,and has important applications in the fields of medical health,smart home,aerospace and military.However,the hybrid brain-computer interface system based on the fusion of EEG and EMG signals greatly limits the multiplexing of different data and cross-individual tasks because of the non-stationarity of EEG signals and the individual differences of EMG signals.In addition,how to integrate the advantages of the two to achieve a more intelligent human-computer interaction method efficiently is the key problem and challenge faced by the hybrid braincomputer interface.Aiming at the above key issues and challenges,the main research work of this paper is as follows:First of all,in view of the different probability distributions of cross-domain EMG data samples caused by interference factors such as electrode position offset,muscle fatigue,and individual differences in EMG signals,as well as the poor generalization ability of traditional transfer learning methods and the easy occurrence of negative transfer,etc.An EMG signal transfer learning method based on improved Easy TL is proposed,which not only selects the features that are beneficial to crossdomain transfer at the feature level,but also fundamentally solves the difference in the probability distribution of samples from different fields in the feature space.The experimental results show that the proposed method greatly reduces the distribution differences of data samples of cross-domain EMG signals,and solves the problem of negative transfer effectively.Secondly,aiming at the non-stationarity problem caused by different measurement states of EEG signals and the limitation of artificial feature extraction on the deep features of EEG signals,the EEGNet-CORAL deep adaptive network is proposed.In the methods,the deep neural network is used to perform deep features on EEG signals.At the same time,using covariance matrix alignment to perform domain adaptation on the extracted deep features,reducing the probability distribution difference of features between different domains.The experimental results show that the proposed method reduces the distribution differences of crossdomain EEG data samples and improves the robustness and generalization ability of the source domain model on the target domain.Finally,for solving the poor fusion effect caused by the inconsistent contributions of EEG and EMG signals,the unsatisfactory feature layer fusion and decision layer fusion methods,and the distribution differences of EEG and EMG signals across individuals,a EEG-EMG feature fusion method based on GA-SVM and cross-individual fusion strategy of first fusion and then transfer was proposed.The experimental results show that the proposed method not only realizes the optimal combination of EEG-EMG features,but also reduces the probability distribution difference between cross-individual EEG signals and EMG signals.Experiments further confirm the superiority of the cross-individual fusion strategy of fusion first and then transfer.