An Unsupervised Deep-Transfer-Learning-Based Motor Imagery EEG Classification Scheme for Brain-Computer Interface.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) are rapidly advancing, with electroencephalography (EEG) being a key non-invasive technique.
• EEG signals are non-stationary, leading to data distribution variations across time and subjects, which hinders BCI performance.
• Existing BCI systems face limitations due to these signal variations, restricting their practical applications.

Purpose of the Study:

• To propose an unsupervised deep transfer learning method for motor imagery EEG signal classification.
• To address the challenges posed by non-stationary EEG data in BCI systems.
• To enhance the robustness and applicability of BCI technology.

Main Methods:

• Utilized transfer learning for motor imagery EEG signal classification.
• Employed Euclidean space data alignment (EA) to align covariance matrices of source and target EEG data.
• Applied Common Spatial Pattern (CSP) for feature extraction from aligned data.
• Implemented a deep convolutional neural network (CNN) for final EEG classification.

Main Results:

• The proposed deep transfer learning method demonstrated effectiveness in classifying motor imagery EEG signals.
• Experimental results on public EEG datasets validated the method's performance.
• The approach successfully addressed data distribution discrepancies inherent in EEG signals.

Conclusions:

• The developed unsupervised deep transfer learning method offers a promising solution for improving BCI system performance.
• The technique effectively handles non-stationary EEG data, enhancing classification accuracy.
• This study contributes to advancing the practical application of BCI technology through improved signal processing and machine learning.