Motor Imagery EEG Classification Based on Transfer Learning and Multi-Scale Convolution Network.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Accurate electroencephalography (EEG) signal recognition is crucial for brain-computer interface (BCI) system success.
• Individual EEG signal variability and limited data pose challenges for classification and recognition.

Purpose of the Study:

• To address individual EEG differences and data scarcity in BCI applications.
• To develop and evaluate an attention mechanism-based multi-scale convolution network integrated with transfer learning for motor imagery EEG signal analysis.

Main Methods:

• A novel dual-channel attention module migration alignment with convolution neural network (MS-AFM) was designed.
• Transfer learning data alignment algorithm was introduced to enhance motor imagery EEG signal analysis.
• The BCI Competition IV dataset 2a was utilized for model verification.

Main Results:

• The proposed MS-AFM model demonstrated improved classification recognition rates when incorporating the alignment algorithm and adaptive transfer learning.
• An average classification recognition rate of 86.03% was achieved across nine subjects.
• The integration of transfer learning effectively mitigated challenges associated with individual EEG signal differences and data limitations.

Conclusions:

• The developed MS-AFM model with transfer learning significantly enhances the accuracy of EEG signal recognition for BCI systems.
• This approach offers a promising solution for overcoming common limitations in BCI research, particularly concerning data variability and quantity.
• The findings highlight the potential of advanced deep learning architectures and transfer learning strategies in advancing BCI technology.