Emotion Recognition Method Based on EEG Signals: 4D-CapsBLnet Model Integrating Spatio-Temporal Features

Emotion recognition technology based on electroencephalogram (EEG) signals has become a focal point for experts and scholars globally. To improve the accuracy of machine understanding of human emotions, this paper characterizes the multidimensional information features in EEG signals, and solves the problem of 4-dimensional fusion of temporal, frequency and 2D spatial features of EEG signals, and we present a novel method named 4D-CapsBLnet. First, we transformed the multidimensional features from various channels into 4D structures for training the deep learning model. Second, we utilized Capsule Networks (CapsNet) to capture spatial and frequency information from each temporal slice of 4D inputs, while Bidirectional Long Short- Term Memory Networks (Bi-LSTM) was employed to capture temporal dependencies. To comprehensively assess the model's performance, we conducted a five-fold cross-validation with a maximum epoch limit and early stopping strategy. The average accuracy and standard deviation on the SEED, SEED-IV, and SEED-V datasets are 95.06% ± 1.65%, 88.47% ± 1.74%, and 88.23% ± 1.12%, respectively. Compared to similar research methods, the proposed model in this paper demonstrates robustness and superior performance in emotion recognition tasks.