Abstract: The channels and time segments of electroencephalogram (EEG) signals in the process of emotion generation contained rich emotional information, and different time segments held varying importance in emotion recognition. The challenge was how to capture key features and highlight key time segment information, we proposed a  multi-|dimensional feature  emotion recognition method based on EEG. This method fully utilized  frequency, spatial,  temporal characteristics, and attention information of physiological signals. By constructing a four-dimensional feature matrix combined with a depthwise separable network and embedding a convolutional sliding window to adaptively extract the spatial-frequency features of EEG signals. Meanwhile, a multi-head attention mechanism was integrated into the temporal convolutional neural network to highlight  important time series information and achieve emotion recognition. The accuracy of wake-up and potency of the proposed method on the DEAP dataset is 97.49% and 97.36%, respectively, and the accuracy of the method on the SEED dataset is 96.60%, which is about 3% higher than that of the mainstream method. The experimental results verify the superiority of the model in physiological signal emotion recognition.

Key words: emotion recognition, electroencephalogram signal, temporal convolutional network, multi-head attention mechanism