Feature Extraction And Classification Algorithm For EEG-based Emotion Recognition

Emotion is a manifestation of human intelligence,and with the development of deep learning and artificial intelligence technologies,using EEG signals to identify human emotions has become a hot research direction in the field of bioinformatics recognition.EEG emotion signal is a kind of nonlinear and non-smooth continuous random time series signal,and traditional machine learning methods are no longer sufficient to extract more perfect and rich EEG emotion features with higher-order semantic information.Using deep learning to automatically extract pure data-driven EEG emotional features from raw EEG data can compensate for the shortage of traditional manual feature extraction.Therefore,in this paper,we conduct a study on EEG feature extraction and classification algorithms for emotion recognition based on deep learning methods and combining a priori knowledge of EEG emotion signals.The specific research work is as follows:(1)Aiming at the problem that the spatial location information of brain electrodes is not fully considered in the process of EEG emotion recognition,this paper proposes an EEG emotion feature extraction algorithm that fuses spatial and frequency information.The feature extraction method mainly consists of three parts: EEG time-frequency domain conversion,brain electrode spatial information mapping and null-frequency feature matrix construction.The frequency information of multi-channel EEG emotion signal is extracted by Fourier transform,calculation of power spectral density and differential entropy information of each frequency band;the electrode mapping matrix is constructed according to the principle of the nearest position,and then it is stitched into a null-frequency feature matrix with a priori knowledge from top to bottom and from left to right.Through experimental verification,this EEG emotion feature can promote the accuracy of EEG signal-based emotion recognition.(2)To address the problems of large data volume,information overload and the consistent importance of all EEG channels for multi-channel EEG emotion recognition,this paper proposes an attention mechanism-based EEG physical channel enhancement method.The method first automatically extracts the time-frequency-like feature maps of EEG emotional signals by scaling convolutional layers,and then reassigns importance weights to each EEG channel by the EEG channel enhancement module to enhance the EEG channels relevant to the emotional recognition task,while suppressing the EEG channels that are not relevant to them.Finally,the validity and feasibility of this EEG channel enhancement method are demonstrated by experimental validation with the visualization of brain topography maps of different emotional dimensions,and the accuracy of EEG emotion recognition is also improved.(3)This paper proposes an EEG emotion recognition network,SFMS-Net,which combines null-frequency features with a priori knowledge and augmented time-frequency features to improve the accuracy of multi-channel non-specific EEG emotion recognition.The accuracy of multi-channel non-specific EEG emotion recognition is effectively improved.Finally,through validation on the DEAP EEG emotion dataset,the recognition accuracies in the three emotion description dimensions of Valance,Arousal and Dominance reached 71.41%,71.49% and 71.65%,respectively,compared with the latest cutting-edge non-specific EEG emotion recognition related research in 2021.Their emotion recognition accuracies improved by 0.28%,1.5% and 0.87%,respectively.The research work in this paper combines traditional machine learning manual feature design with automatic feature extraction using deep learning,and proposes an EEG emotion feature extraction algorithm,an EEG channel enhancement method and an emotion recognition network,which effectively improves the accuracy of EEG emotion recognition.It is important for promoting research in the field of brain-machine emotion interaction and bioinformation recognition.