EEG Emotion Recognition Method Based on a Hybrid Perception Training Strategy

Emotion recognition technology based on electroencephalogram (EEG) has attracted much attention in brain activity analysis due to its millisecond-level temporal resolution. However, bottleneck problems such as significant differences in neural responses between individuals and complex spatiotemporal coupling characteristics of signals restrict the effectiveness of practical applications. This study innovatively proposed a hybrid perception-driven auxiliary enhanced spatiotemporal network (AETSN), whose core architecture adopts a dual-path collaborative mechanism: on the one hand, the electrode topological relationship is modeled through a graph convolutional network (GCN) to analyze the functional connection characteristics of brain regions, and on the other hand, the self-attention Transformer module is introduced to deeply mine the temporal dynamic patterns related to emotions. In order to further improve the generalization ability across subjects, a multimodal auxiliary training strategy is designed - the potential emotion representation of peripheral physiological signals (ECG/EDA) is extracted through pre-training auxiliary networks, and the feature alignment loss function is used to realize the knowledge transfer of the main and auxiliary networks. Experiments based on the DEAP benchmark dataset show that the model achieved an accuracy of 62.58% and 59.01% in the arousal and valence binary classification tasks, respectively, which is an average improvement of 2.3%-4.5% over classic deep models (such as EEGNet, MASA-TCN, etc.). The ablation experiment further verified the effectiveness of the spatial-temporal feature fusion module and the multi-source data collaborative training mechanism, providing a new idea for lightweight cross-subject emotion computing.