Jump motion intention recognition and brain activity analysis based on EEG signals and Vision Transformer model

The lower limb exoskeleton can help the human jump movement and expand jump ability. However, recognizing jump motion intention promptly and accurately is a challenge. The electroencephalography (EEG) signal has the potential to recognize jump motion early due to accompanying motor intention generation. This paper proposes a method based on Vision Transformer to extract spatial-spectral-temporal information from multi-channel EEG signals to recognize jump motion intentions including pre-jump, off-ground, and post-jump. Artifacts of EEG signals are removed by filtering and independent component analysis. The movement-related cortical potential, time-frequency, source localization, and functional connectivity analyses are performed to analyze the changes in brain activity during jump, which can explore motion control mechanism of the brain and provide physiological basis for the construction of the recognition method. The EEG features in time domain and frequency domain are reduced dimensionally by channel, feature, and frequency band selections. The proposed model based on multi-head self-attention architecture extracts spatial information from sequence composed of multi-channel temporal and spectral fusion features. The recognition performance of the proposed model outperforms those of comparison models, and the average recognition accuracy and kappa coefficient are 88.797% and 0.8325, respectively. The classification performance of temporal and spectral fusion features is higher than that of temporal features, especially spectral features. The brain networks constructed by attention coefficients focus on the connectivity differences among tasks and cannot reflect the physiological significance of EEG signals. Finally, the proposed method is validated on the open access BCI competition IV Dataset 2a of EEG signals.