Motor imagery classification based on dynamic multi-scale convolution and multi-head temporal attention.
10.7507/1001-5515.202408051
- Author:
Nan XIAO
1
;
Ming'ai LI
1
Author Information
1. College of Information Technology, Beijing University of Technology, Beijing 100124, P. R. China.
- Publication Type:Journal Article
- Keywords:
Dynamic convolution network;
Motor imagery electroencephalogram;
Multi-scale convolutional network;
Self-attention mechanisms
- MeSH:
Electroencephalography/methods*;
Humans;
Neural Networks, Computer;
Brain-Computer Interfaces;
Attention;
Signal Processing, Computer-Assisted;
Imagination/physiology*;
Algorithms
- From:
Journal of Biomedical Engineering
2025;42(4):678-685
- CountryChina
- Language:Chinese
-
Abstract:
Convolutional neural networks (CNNs) are renowned for their excellent representation learning capabilities and have become a mainstream model for motor imagery based electroencephalogram (MI-EEG) signal classification. However, MI-EEG exhibits strong inter-individual variability, which may lead to a decline in classification performance. To address this issue, this paper proposes a classification model based on dynamic multi-scale CNN and multi-head temporal attention (DMSCMHTA). The model first applies multi-band filtering to the raw MI-EEG signals and inputs the results into the feature extraction module. Then, it uses a dynamic multi-scale CNN to capture temporal features while adjusting attention weights, followed by spatial convolution to extract spatiotemporal feature sequences. Next, the model further optimizes temporal correlations through time dimensionality reduction and a multi-head attention mechanism to generate more discriminative features. Finally, MI classification is completed under the supervision of cross-entropy loss and center loss. Experiments show that the proposed model achieves average accuracies of 80.32% and 90.81% on BCI Competition IV datasets 2a and 2b, respectively. The results indicate that DMSCMHTA can adaptively extract personalized spatiotemporal features and outperforms current mainstream methods.
