Emotional EEG Recognition Based on Multi-scale Spatio-Temporal Attentional Convolution Network and Domain Adaptation

Emotion recognition based on deep learning has always been a hot topic in the brain-computer interface field. Neuroscientific studies have shown that EEG signals contain information in different time-frequency components, and exhibit different activation patterns in different spatial regions. Moreover, in practical applications, individual data is often restricted, leading to models overly relying on a limited number of samples. To fully represent EEG signals and improve model generalization ability, this paper proposes a method for emotional EEG recognition based on multi-scale spatio-temporal attentional convolution network and domain adaptation (MSSTACN-DA). It extracts multi-scale temporal features through multiple parallel dynamic spectrum-temporal convolutions, learns local and global spatial information of EEG signals through asymmetric spatial convolutions, and integrates temporal and spatial attention to focus on important information of different dimensions. Furthermore, we design a domain adaptation mechanism, where the model is first pretrained and then fine-tuned with a small amount of target domain data to improve performance. The proposed method is validated on SEED dataset, and the results demonstrate that after domain adaptation, the accuracy reaches 99.56%, confirming the feasibility and effectiveness of the proposed method.