Human Intention Recognition Based On EEG Signals

The EEG-based intention recognition algorithm analyzes EEG signals to determine human intention activities,which can provide limb replacement technology for disabled people,assist paralyzed patients in recovering their motor abilities,improve the interaction efficiency of intelligent devices,and have broad application prospects in medical,military,and entertainment fields.However,there are still three problems with the existing intention recognition algorithms:(1)Large individual data distribution differences lead to decreased accuracy of the intention recognition model;(2)Inaccessible existing user data leads to difficulty in cross-subject recognition;(3)Non-stationary fluctuations in EEG data distribution lead to decreased accuracy of customized models.In response to these key issues,this thesis has completed the following work:For the cross-subject intention recognition scenario,this thesis proposes a Bi-classifier Similarity Clustering Domain Adaptation(SC-DA)algorithm based on similarity clustering,which addresses the problem that existing domain adaptation algorithms ignore the impact of similar intention activity samples on unbiased feature learning during the transfer initial stage,resulting in decreased accuracy of the intention recognition model.SC-DA introduces a label consistency strategy on the basis of the bi-classifier adversarial domain adaptation method,which makes the predicted labels consistent between samples with high local similarity,so that the model can accurately recognize similar intention activity samples in the early stage of domain adaptation and form a clearer decision boundary for the intention recognition.Experimental results on the BCI-IV 2a and BCI-IV 2b public datasets show that compared with existing algorithms,the proposed SC-DA algorithm improves the average accuracy by 3.2% and 1.4%,respectively.For the privacy-protected cross-subject intention recognition scenario,this thesis proposes a Source-free Intention Recognition(SF-IR)algorithm based on domain-invariant features,which addresses the problem that existing source-free domain adaptation algorithms ignore the impact of non-stationarity of EEG signals on the feature extractor in intention recognition,resulting in bias in the feature extractor during model initialization.SF-IR uses a two-stage framework,where the first stage uses the original data to generalize the model,making the model focus more on event-related potential(ERP)information strongly related to intention task,and the second stage uses the target individual’s data for cluster fine-tuning.Experimental results on the BCI-IV 2a and BCI-IV 2b public datasets show that compared with existing algorithms,the proposed SF-IR algorithm improves the average accuracy by 3.2% and 1.2%,respectively.For the cross-session intention recognition scenario for practical user customization,this thesis proposes a Weighted Adversarial Domain Adaptation(WADA)algorithm,which addresses the problem that existing algorithms only consider the alignment of the overall distribution and ignore the inconsistent transfer value of features in different regions in cross-session tasks,resulting in negative transfer of the model.WADA optimizes the transfer features based on information entropy to achieve a finer-grained feature space alignment.Experimental results on the BCI-IV 2a public dataset and the collected dataset show that compared with existing algorithms,the proposed WADA algorithm improves the average accuracy by 2.1% and 3.2%,respectively.