Temporal Fusion Dynamically Separable Graph Convolutional Network for EEG Motion Intention Decoding Based on Source Information Extraction

The electroencephalogram (EEG) signals acquired through scalp electrodes are susceptible to the volume conduction effect, which introduces a considerable degree of redundant information and consequently hinders decoding accuracy. In response to this challenge, we introduce a novel methodology that incorporates the Laplacian matrix of a graph to construct a source information extraction module, effectively mitigating the impact of volume conduction in electrodes. This innovation culminates in the development of the source information extract-based multiscale temporal fusion dynamic separable graph convolution network (SIE-TFDSGCN). The proposed method employs varying lengths of time convolution for fusion, thereby extracting temporal features inherent in EEG signals. Furthermore, a dynamic separable graph convolution (DSGCN) structure is introduced to extract spatial features, concurrently further reducing the influence of volume conduction. Performance evaluation on the BCI Competition IV 2a public dataset yielded notable results, with an average accuracy of 82.48% and a kappa value of 76.65. Comparative analysis against state-of-the-art classification methods underscores the significant enhancement in classification accuracy achieved by our approach. Expanding upon these findings, we leverage the developed methodology to design a brain-computer interface (BCI) smart home online control system. This system exhibits a high degree of reliability, enabling effective real-time control for smart homes. Our contributions mark a substantial advancement in EEG signal processing, with implications for both neuroscientific research and practical applications in smart home technology.