Research On EEG Signal Decoding Method For Joint Quantification Of Sample And Feature Quality

With the rapid development of brain science,Brain-computer interface technology is used in a wide range of fields.As one of the main physiological signals,the study of EEG signals has received a lot of attention.The characteristics of the EEG signal can easily lead to different contributions of the extracted samples and features of the EEG signal to pattern recognition.However,existing studies either do not consider the importance of the sample and the characteristics or consider only one of them.The discovery of implicit information in EEG signals is also a jewel in the field of brain science.With the in-depth study of EEG signals,we found that the signal can characterize some physical information,but the current research is still poorly understood in this area,and there are problems such as weak interpretability.Based on the current problems in the field of EEG signals,this thesis proposes improvements in two aspects.1)Addressing the problem of inconsistent sample and feature contributions in EEG signals leading to reduced accuracy in pattern recognition tasks.In this thesis,we propose a machine learning model for the joint sample and feature importance assessment(JSFA).The model assigns a weight to each EEG sample and feature based on self-paced learning and adaptive learning of feature weights.The weight of the sample is obtained by the loss function of the sample during training,the larger the value of the loss function,the lower the weight of the sample.The weights of the features are obtained by adaptive learning,by embedding the feature weights into a least squares regression model and solving for the weights of each feature at the end of training.A theoretical proof of the sample and feature quantification method was performed.The model is validated through a series of experiments,starting with a visual demonstration of sample and feature quality quantification on a synthetic dataset.Then the cross-time sentiment recognition and fatigue detection tasks were completed on the SEED dataset.Experimental results show that our proposed model achieves better recognition accuracy on both EEG signal decoding tasks.2)To address the problem of small samples in EEG signals,this thesis proposes to use only a small number of labeled samples and a large number of unlabeled samples to complete the EEG signal decoding task,extending the JSFA to semi-supervised learning.On the one hand,by incorporating unlabeled samples into model training,the increase in the number of samples can better capture the basic characteristics of EEG data;on the other hand,the label matrix of unlabeled samples is jointly optimized with other variables,which is beneficial to model iteration and training.Experiments on the SEED-IV dataset showed that the semi-supervised joint sample and feature importance evaluation(s JSFE)improved the recognition accuracy compared to the semi-supervised least squares regression model,and the average accuracy was better than JSFA model.In addition,the sample importance vector was projected to the two-dimensional plane by t-SNE,and the conclusion of " samples-label inconsistency" was obtained,which was manifested by the overlapping data distributions of sad and fear trials,which helped to optimize the experimental paradigm of EEG data collection and visualize the frequency band maps and brain area maps of feature importance.From a Data-driven perspective,Delta and Gamma frequency bands contribute more to emotion recognition,and the brain regions most associated with the occurrence of emotion effects are temporal and prefrontal lobes.