Study On Eeg Feature Transfer Learning Method For Emotion Recognition

Emotion recognition is a key issue in the field of emotion brain-computer interface.It has wide impacts on road traffic safety,human-computer interaction,medical health and so forth.In recent years,physiological signals,especially EEG signals,can effectively reflect a person's emotional state.Besides,because of their difficulty in camouflage,strong real-time differences,and high accuracy,EEG-based emotion recognition has become the main research content in emotion recognition.However,due to the non-stationarity of EEG signals,there is difference in data distribution between EEG signals from various tested subjects,which leads to limitations of traditional emotion recognition modeling methods regarding to generalization and complexity.Transfer learning is widely used because it can better solve the problem of data distribution differences.Unfortunately,most of the current research does not success in processing the data effectively.For this reason,an EEG Feature Transfer Learning Method for emotion recognition is presented in this report,and universal emotion induction website platform is designed and developed.The specific work of this paper is demonstrated as follows:Firstly,this paper introduces the background and the current research status at home and abroad of emotion recognition based on EEG.Then the research purpose and significance are explained.This paper will also introduce the characteristics of EEG signals and emotion model classification,analyze the basic principles of emotion recognition and transfer learning.Secondly,in order to obtain the correlated information among data,the improved Max-Relevance and Min-Redundancy algorithm can effectively select the features of source domain and target domain from the level of feature similarity and correlation.In order to select an appropriate source domain for migration among multi-source domains,a multi-source domain adaptive selection algorithm is presented,and the clustering algorithm is used to adaptively select and assign weights to the multi-source domain from the domain-domain difference and sample-domain difference level;Further verify the effectiveness of the proposed method on the emotional EEG SEED data setThirdly,the improved Max-Relevance and Min-Redundancy algorithm and the multi-source adaptive selection algorithm are introduced into the manifold embedded distribution alignment algorithm,and the improved manifold embedded distribution alignment algorithm is presented.While retaining the advantages of traditional algorithms to avoid feature distortion and quantitatively estimate the importance of marginal distribution and conditional distribution,through the process of eliminating invalid features between each source domain and target domain based on the improved Max-Relevance and Min-Redundancy algorithm,obtaining the source domain suitable for migration according to a multi-source domain adaptive selection algorithm,collecting the emotion recognition results on the basis of sending the source domain weights and corresponding recognition results to the decision-level fusion model,it will have better multi-source domain migration capabilities,while reducing the computational complexity,and effectively reducing the phenomenon of negative migration,improving the emotion recognition accuracy and robustness;Experimental comparative analysis on the emotional EEG SEED data set is conducted,verifying the effectiveness of the improved manifold embedded distribution alignment algorithm.Finally,combined with the existing emotion-inducing methods,an emotion-induced website platform designed to induce changes in the emotional state of the subjects was developed,including the design and implementation of front-end and back-end interfaces and functions,which will assist improving the convenience of following researchers' experiments;Through the data collected in this platform and comparative analysis made from designed experiment,the practicability of the emotion-induced website platform is proved.At the same time,the effectiveness of the improved manifold embedded distribution alignment proposed in this paper is verified.