A Temporal–Spatial Embedding and Dynamic Aggregation Network With Adaptive Weighting Spectrum for EEG Motion Intention Recognition

Motor imagery (MI) intention recognition using electroencephalogram (EEG) has emerged as a research hotspot in the field of brain-computer interfaces (BCIs). However, most current methods rarely consider the temporal-spatial dynamic association across different dimensions, leading to insufficient temporal-spatial feature extraction. Furthermore, existing methods tend to learn spectral features empirically, which may fail to identify the most discriminative spectral features. To address these limitations, we propose a novel temporal-spatial embedding and dynamic aggregation network with an adaptive weighting spectrum (TSE-DA-AWS) for MI-EEG recognition. Specifically, we first design a dual-branch TSE block to extract temporal and spatial dependencies of EEG signals, capturing temporal relationships between different time positions and uncovering meaningful spatial interactions among EEG channels. Second, we investigate a temporal-spatial DA block to integrate the associations of temporal-spatial features with time by a dynamic graph aggregation mechanism. Finally, we utilize an AWS block to learn an optimal weight for global spectral information, which can automatically acquire discriminative spectral features. Competitive experimental results on three public datasets validate the efficacy of the proposed method, indicating that our proposed method achieves superior classification performance of 83.62% on the BCI IV 2a dataset, 87.32% on the BCI IV 2b dataset, and 95.63% on the HGD dataset.