Research On Motion Imagery Decoding Method Based On EEG Signal Credibility Index

As a technology that enables communication between the brain and external devices,Motor Imagery Brain-Computer Interface(MI-BCI)has attracted wide attention.It generates brain signals that correspond to control commands without requiring any external stimulation or actual movement from the user.MI-BCI has gradually been applied in the field of rehabilitation medicine.Decoding the brain signals is the core of MI-BCI,and how to train a good classification model using collected EEG data is crucial.Current research assumes that the labels obtained during the experiment are absolutely reliable,ignoring the high degree of uncontrollability of spontaneous EEG signals during motor imagery.This may lead to mislabeling of the data.Furthermore,existing methods often fail to take into account the multiple features of MI-BCI signals,making it difficult to further improve performance.Therefore,this thesis proposes a credibility evaluation method for MI-BCI signals and integrates it into a spatiotemporal attention-based decoding method.A transfer learning strategy is also introduced to improve the generalization ability of the method.The specific research work includes:1.Regarding the problem of difficult quantification of label reliability,this thesis proposes four EEG signal reliability evaluation indicators with different information representations,namely: band power ratio indicator,nonlinear dynamical indicator,brain laterality indicator,and brain connectivity indicator.Then,the significant differences of these indicators were verified in different task states.Then,by comprehensively using these four indicators,the original dataset was divided into two groups according to the level of reliability,and classic machine learning algorithms were used to decode the two groups of data sets separately.The experimental results showed that under classic machine learning algorithms,out of 109 subjects,75 subjects achieved higher classification accuracy in the group with higher reliability evaluation.2.To address the problem of low accuracy caused by abnormal labels,this thesis proposes a method for calculating sample weights based on the work of the first study.By assigning lower weights to samples with lower reliability,the model focuses less on abnormal samples,thus reducing the negative impact of abnormal labels.To enable the classifier to focus on the frequency,spatial,and temporal information of the EEG signals,this thesis draws inspiration from the filter group common spatial pattern and proposes a Transformer model based on spatiotemporal-frequency attention.This model is combined with sample weights,and a filter group Riemannian geometry feature method is also used.The results show that by adding the sample weight mechanism,the two methods’ average accuracy for four-class classification of motor imagery increased from 81.8% and 73.9% to 83.0% and 75.8%,respectively.3.In order to address the problem of large individual differences in motor imagery EEG signals and improve the robustness of the decoding method,this study optimized the decoding model using domain adversarial training strategy(DANN),which improved the cross-subject classification accuracy by 25.9% under the optimal transfer scheme.This study also adopted different data selection schemes based on signal reliability evaluation indicators,and the results showed that data with relatively low signal reliability would have a certain adverse effect on the transfer process.