Meta-Learning-Based EEG Signal Processing Techniques And Their Application In Unmanned Vehicle Control

As an emerging interactive technology,Brain-Computer Interface(BCI)provides a new way for human-computer interaction and holds great potential in fields such as rehabilitation medicine,military,and personal entertainment.EEG signal classification is crucial in BCI applications.Currently,deep learning algorithms have been widely used for BCI signal classification.However,these models require a large amount of labeled EEG data for training and have poor generalization capability,limiting the widespread application of BCI.In this thesis,we investigate the use of meta-learning-based EEG signal processing techniques and propose an application for controlling autonomous vehicles using motor imagery EEG signals.The specific research content includes:Firstly,due to the differences in EEG signal characteristics among different users and the non-stationarity of the same user’s EEG signals,traditional EEG signal classification models have poor generalization ability.This thesis introduces multi-source domain adaptation(MSDA)based on Model-Agnostic Meta-Learning(MAML)and proposes an EEG signal classification method based on both.This method first regards different users as different source domains,randomly pairs them into task pairs as the input of the inner model,and then calculates the loss function using the domain differences between each task,calculates the average loss after multiple iterations,updates the general model parameters of the outer layer until the loss function converges.This method can better utilize the meta-knowledge learned on different domains,enabling the trained model to quickly adapt to the EEG signal classification of other subjects after finetuning,reducing the amount of required training data.Based on a publicly available EEG dataset,the proposed method is compared with methods based solely on Convolutional Neural Network(CNN),MAML,and Prototypical Network for performance analysis.The results show that at an 85%classification accuracy,the proposed method can reduce the required training sample size by 50%compared to the CNN-based method.Additionally,the classification accuracy is 5%higher than other metalearning methods under the same training sample size,proving the effectiveness of the proposed method in cross-subject scenarios.Secondly,based on the proposed EEG signal classification method,this thesis designs and implements a real-world BCI-controlled unmanned vehicle system,which includes EEG signal acquisition and processing,communication and video feedback,and unmanned vehicle control modules,solving the problem of cross-subject BCI-controlled unmanned vehicle applications.In the system construction process,the thesis adopts the principle of low coupling and high cohesion for the overall system,and modularizes the design,enhancing the scalability of the BCI-controlled unmanned vehicle and achieving a modular,portable,and highly integrated BCI-controlled unmanned vehicle system.The accuracy and latency of BCI-controlled unmanned vehicle control were tested in a real environment.The results show that the average control accuracy of the BCI-controlled unmanned vehicle application built in this thesis can reach over 80%,which can accurately,stably,and quickly realize EEG signal processing and unmanned vehicle control.In summary,this thesis focuses on cross-subject EEG signal classification,develops a BCI-controlled unmanned vehicle application,and tests it in a real environment,providing guidance for the promotion of cross-subject BCI-controlled unmanned vehicle applications.