| 刊名 | Agricultural Biotechnology |
| 作者 | Zhikang YIN1, Chunjiang SHUAI2* |
| 作者单位 | 1.School of Physics and Telecommunication Engineering, Shaanxi University of Technology, Hanzhong 723000, China; Traien Institute of Technology, Shaanxi University of Technology, Hanzhong 723000, China |
| DOI | DOI:10.19759/j.cnki.2164-4993.2025.06.018 |
| 年份 | 2025 |
| 刊期 | 6 |
| 页码 | 88-95,99 |
| 关键词 | Deep learning; Brain-computer interface (BCI); Convolutional neural network (CNN); Electroencephalogram (EEG); Motor imagery (MI) |
| 摘要 | Deep learning methods have been widely applied in motor imagery (MI)-based brain-computer interfaces (BCI) for decoding electroencephalogram (EEG) signals. High temporal resolution and asymmetric spatial activation are fundamental properties of EEG during MI processes. However, due to the limited receptive field of convolutional kernels, traditional convolutional neural networks (CNNs) often focus only on local features, and are insufficient to cover neural processes across different frequency bands and duration scales. This limitation hinders the effective characterization of rhythmic activity changes in MI-EEG signals over time. Additionally, MI-EEG signals exhibit significant asymmetric activation between the left and right hemispheres. Traditional spatial feature extraction methods overlook the interaction between global and local regions at the spatial scale of EEG signals, resulting in inadequate spatial representation and ultimately limiting decoding accuracy. To address these limitations, in this study, a novel deep learning network that integrates multi-modal temporal features with spatially asymmetric feature modeling was proposed. The network first extracts multi-modal temporal information from EEG data channels, and then captures global and hemispheric spatial features in the spatial dimension and fuses them through an advanced fusion layer. Global dependencies are captured using a self-attention module, and a multi-scale convolutional fusion module is introduced to explore the relationships between the two types of temporal features. The fused features are classified through a classification layer to accomplish motor imagery task classification. To mitigate the issue of limited sample size, a data augmentation strategy based on signal segmentation and recombination is designed. Experimental results on the BCI Competition IV-2a (bbic-IV-2a) and BCI Competition IV-2b (bbic-IV-2a) datasets demonstrated that the proposed method achieved superior accuracy in multi-class motor imagery classification compared with existing models. On the BCI-IV-2a dataset, it attained an average classification accuracy of 84.36%, while also showing strong performance on the binary classification BCI-IV-2b dataset. These outcomes validate the capability of the proposed network to enhance MI-EEG classification accuracy. |
