Research On Cross-Subject And Cross-Dataset Classification Method Of Event-Related Potential Brain-Computer Interface

Brain-computer interface(BCI)based on event related potential(ERP)has important applications in the fields of neurorehabilitation and human-computer interaction.However,due to the large subject differences of ERP and the different collection conditions of different datasets,the ERP-BCI classification model based on machine learning has poor transferability between different subjects and different datasets.Therefore,it is difficult to horizontally compare different ERP-BCI studies using different identification methods.These problems increase the cost of model training in ERP-BCI applications.This thesis has carried out the following work on this issue:First of all,this thesis selects three typical ERP-BCI datasets,which have different paradigms.Dataset I is a fast sequence flash experiment,and dataset II and dataset III are improved P300-Speller experiments.A unified data preprocessing and feature extraction were performed on the three datasets,the ERP of different datasets was extracted using superimposed average,and the ERP time domain feature analysis was performed.By comparing the characteristics of target stimulus ERP and non-target stimulus ERP,the different responses of subjects to target stimulus and non-target stimulus can be clearly distinguished.By comparing the ERP characteristics and brain topographic maps of all lead target stimulations in the three datasets,it is found that the target stimulation ERP waveforms of the three datasets have different latency and amplitude,but the waveform changes are similar.Secondly,this thesis selects classic machine learning methods and deep learning methods for different ERP-BCI datasets to conduct research on intra-subject and crosssubject identification methods.This thesis selects four classic classification algorithms:linear discriminant analysis(LDA),stepwise linear discriminant analysis(SWLDA),naive bayesian model(NBM),extreme learning machine(ELM),and optimizes the network structure parameters of the EEGNet algorithm for cross-subject recognition and screening,and finally determines the learning rate of 0.02,batch size for 200,the dynamic reduction learning rate is 0.99.This paper uses the area under curve(AUC)of the receiver operating characteristic(ROC)as an index to evaluate the classification algorithm,and simultaneously uses sensitivity and specificity to simultaneously evaluate the performance of the classifier.In the study of classification and identification within subjects,the LDA classification performance is the best in the dataset I,with an ROC-AUC value of 0.931±0.033,the SWLDA classification performance in the dataset II is the best,with an ROC-AUC value of 0.920±0.041,and the SWLDA classification performance in the dataset III is the best.The ROC-AUC value is 0.926±0.033.When cross-subject classification and recognition,the ROCAUC values of EEGNet algorithm in dataset I,dataset II and dataset III are 0.900±0.029,0.896±0.037 and 0.897±0.029.Among the four classic classification algorithms,the LDA classification algorithm has the best classification performance,with a separability ROC-AUC value of 0.863±0.040.After paired T-test statistical analysis,it is found that EEGNet’s cross-subject classification performance is significantly better than the four classic classification algorithms,which proves the effectiveness of EEGNet algorithm in cross-subject recognition in different paradigms of ERP-BCI.Finally,this thesis conducts a cross-dataset ERP-BCI recognition research,using one-to-one training and testing models and two-to-one training and testing models,respectively.In the one-to-one training test model,the SWLDA in the classic classification algorithm obtains the highest ROC-AUC value of 0.537±0.019,while the EEGNet ROC-AUC value is 0.658±0.032.In the two-to-one training test model,the classic classification algorithm LDA has the best separability ROC-AUC value of0.524±0.012,while the EEGNet best separability ROC-AUC value is 0.598±0.041.Subsequently,the paired T-test statistical analysis of all ROC-AUC results shows that the classification performance of the EEGNet algorithm is significantly better than the four classic classification algorithms,indicating that the deep learning algorithm has certain application potential in the ERP-BCI cross-dataset classification and recognition problem.In conclusion,this dissertation focuses on the classification and recognition of ERP-BCI,and analyzes the cross-subject and cross-dataset classification and recognition performance of classic algorithms and deep learning EEGNet algorithms in different ERP-BCI datasets.By constructing a cross-subject and cross-dataset classification and recognition model,the effectiveness of the EEGNet algorithm in cross-subject and cross-dataset classification is verified.The research results of this thesis provide references for ERP-BCI’s cross-subject and cross-dataset classification and recognition algorithms.