Research On Motor Imagery EEG Processing Based On Domain Adaptation

The emergence of Brain Computer Interfaces(BCIs)has opened up entirely new possibilities for communication between humans and machines.One of the popular topics in this field is Motor Imagery(MI),which refers to the mental simulation or imagination of specific body movements without actually performing physical actions.By monitoring and interpreting the user’s brain activity,Motor Imagery Brain Computer Interfaces(MI-BCIs)enable individuals to control external devices through the power of thought,even in the absence of actual physical movement.Electroencephalographic(EEG)signals are commonly used inputs in BCIs,gaining widespread attention due to their high temporal resolution and non-invasiveness.Decoded through algorithms,EEG signals can be employed to identify users’ cognitive states,intentions,or behaviors.However,due to various influences such as physiological states and environmental factors,there are distributional differences in EEG signals both across individuals and within individuals.This implies that when processing the EEG signals of a current subject,signals from other subjects or historical data cannot be directly borrowed,hindering the analysis,modeling,and data sharing of EEG signals.To overcome this obstacle,Domain Adaptation(DA)is gradually introduced into EEG signal decoding,aiming to adjust the feature representations of different domains to reduce distributional differences in the feature space.This method helps ensure the optimal utilization of existing data,extract crucial features,and thus improve the classification performance of the current subject.Based on DA and focusing on MI-EEG signals,this dissertation revolves around the inherent inter-individual and intra-individual differences present in the signals.The main work contents are as follows:1.In view of the inter-individual differences in MI-EEG signals,in order to alleviate the problems of lack of utilization of class-level information,excessive attention to domain common information and weak feature space constraints,based on the adversarial training method,this dissertation proposes a method named MI-EEG Decoding Domain Adaptation Algorithm Based on Dual Classifiers and Dual Attention.The algorithm combines domain-level information and class-level information present in the data.It achieves feature distribution alignment through an iterative adversarial process by analyzing the output differences between two classifiers.To prevent negative transfer resulting from enforced domain alignment,the algorithm proposes the utilization of non-shared mechanisms to preserve domain-specific information.Additionally,the algorithm introduces a triplet loss to further constrain the target feature space.The algorithm proposed in this dissertation is compared with six mainstream algorithms on a total of eighteen subjects from BCI Competition IV Dataset 2a and BCI Competition IV Dataset 2b.Compared with the best-performing adversarial learningbased EEG decoding algorithms on both datasets,this proposed model achieved better performance on fifteen subjects.2.To address the issue of ignoring inter-sample information in the aforementioned method,this dissertation introduces a method called the MI-EEG Decoding Domain Adaptation Algorithm Based on Comprehensive Information.Building upon the algorithm proposed in Content 1,this enhanced algorithm incorporates sample-feature attention module and graph convolutional module to ensure the utilization of comprehensive information,encompassing domain-level information,class-level information,and inter-sample information during the distribution alignment process.This algorithm fully captures the relationship between samples through the multidimensional weighting method of the sample-feature attention module and the topological structure information extracted by the graph convolution module to ensure all-round exploration of information.The algorithm proposed in this dissertation is compared with the algorithm from Content 1 on a total of eighteen subjects from BCI Competition IV Dataset 2a and BCI Competition IV Dataset 2b.Experimental results show that the proposed algorithm achieves performance improvements of varying degrees for the majority of subjects.3.Convolutional neural networks face certain difficulties in extracting long-range information in features.At the same time,the parameter sharing method of convolution kernels limits the flexibility of information transfer.In contrast,Transformer models leverage their core attention mechanism to effectively model complex dependencies in long sequential data.In view of the above characteristic,this dissertation attempts to solve the individual differences in EEG signals from the perspective of non-adversarial learning and proposes an algorithm named MI-EEG Decoding Domain Adaptation Algorithm Based on Cross-Attention.The algorithm fully exploits the attention mechanism in Transformer to handle long-distance dependencies between features.Additionally,it introduces a cross-domain decoder that enables bidirectional information interaction and knowledge transfer between two domains through a fourbranch approach.The aim of this approach is to better adapt to differences between subjects and enhance the classification performance of the current subject.The algorithm proposed in this dissertation is compared with thirteen other algorithms on a total of twenty-three subjects from BCI Competition IV Dataset 2a,BCI Competition IV Dataset 2b,and BCI Competition III Dataset 2a with different amounts of training data for each subject.Experimental results demonstrate that the proposed algorithm outperforms existing non-adversarial learning-based EEG decoding algorithms.Compared with the best-performing algorithms on the three datasets,the proposed model achieves performance improvements of 2.74%,1.72%,and 5.67% in terms of average accuracy,respectively.4.In view of the intra-individual differences in EEG signals,facing the situation of whether the real labels of some target signals are available for the current subject,taking into account the lack of fine-grained category information utilization in the distribution alignment process and the non-personalized problem of spatial constraints,this dissertation introduces an instance-level contrasting prototype analysis method and provides semi-supervised and unsupervised algorithm architectures accordingly.During the process of aligning feature distributions,these frameworks take into account global information and local category knowledge to comprehensively alleviate distribution differences.Simultaneously,in the unsupervised scenario,recognizing the adverse effects of excessive pseudo-label noise,the algorithm evaluates each subject individually from an entropy-based confidence perspective to dynamically adapt the application of target domain category information.The algorithm proposed in this dissertation is compared with eleven algorithms on a total of sixty-three subjects from the Giga DB and BCI Competition IV Dataset 2a.Compared with the best-performing algorithms on the two datasets,the proposed model achieves performance improvements of 8.13% and 4.02% in terms of average accuracy,respectively.In summary,this dissertation proposes new domain adaptation methods from different perspectives for various scenarios to achieve distribution alignment across individuals and within individuals,ensuring the full utilization and extraction of information.The dissertation provides comprehensive solutions for MI-BCIs when confronted with diverse EEG signals and is expected to foster their further development in practical applications.