Emotion Recognition Of EEG Based On Multi-scale Time Window Combination Features

In recent years,with the rapid development of EEG acquisition equipment,a kind of wireless head-mounted single-electrode EEG acquisition equipment that does not need to be coated with conductive glue has emerged.Compared with the cumbersome and complicated collection methods of multi-electrode equipment,this type of device has gradually become an important option for studying EEG emotion recognition due to its ease of use and flexibility.However,there are few related studies on single-electrode EEG devices,and the accuracy of emotion classification,the generalization ability of emotion recognition model and EEG representation ability still need to be improved.Therefore,the paper proposes an EEG emotion recognition method based on multi-scale time window combination features,which provides a new feasible solution for the research on emotion recognition of single-electrode EEG signals.The main work and content of this paper are as follows:(1)EEG data processing.Aiming at the problem of low recognition accuracy caused by the high noise content and insufficient effective features in the original EEG signal.The paper first uses discrete wavelet transform for denoising,and then uses wavelet transform to decompose the preprocessed EEG signal and extract four frequency bands:Alpha,High Beta,Low Beta and Gamma.Finally,the average value,standard deviation,maximum,minimum,energy and other characteristic attributes of each window are calculated and counted through time windows,which are used as data materials for EEG emotion recognition tasks.(2)Emotional feature extraction and selection.Aiming at the problem that most of the current emotion recognition studies use a single fixed size time window to segment EEG signals and extract a single type and small amount of emotion feature information.The paper proposes a multi-scale time window combination features extraction method,which uses a 1-second time window to segment the EEG signals of three groups of emotion(neutral,positive and negative)labels into multiple sets of overlapping signal segment with 50%.In order to extract as many emotional features and feature information as possible,the 1-second time window is further subdivided into 1/2^k second-wide time windows,and the combination features within each window are extracted separately.And finally,four sentiment recognition classification methods:linear discriminant analysis,support vector machine,K-nearest neighbor and random forest,are used for three-category emotion recognition.(3)Emotion recognition and classification.Aiming at the problems that emotion classification methods are difficult to effectively utilize the time series information in the deep features of EEG signals leading to low recognition accuracy and weak characterization ability.The paper proposes an emotion recognition model,which based on LSTM for the purpose of recognizing and classifying emotion.The model proposed in the paper mainly takes the advantages of LSTM's better performance in solving the problems of long-term dependencies to further extract the time-series information of the EEG signal.The recognition accuracy and classification performance of the experimental results have been greatly improved,and the model has better robustness and representation ability.