EMG-based Multi-User Hand Gesture Classification via Unsupervised Transfer Learning Using Unknown Calibration...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG)-based human-machine interaction (HMI) systems face challenges in generalization and training burden, hindering commercialization.
• Current sEMG gesture classification models often require extensive calibration data, limiting user accessibility.

Purpose of the Study:

• To explore and compare unsupervised transfer learning (TL) algorithms based on convolutional neural networks (CNNs) for sEMG gesture classification.
• To reduce the calibration data requirements and improve the generalization performance of sEMG-based HMI systems.

Main Methods:

• Eight unsupervised TL algorithms using CNNs were evaluated on a dataset of 10 gestures from 35 subjects.
• Performance was compared against traditional classifiers (KNN, LDA, SVM, Random Forest) using minimal calibration data.

Main Results:

• The CORrelation Alignment (CORAL) TL algorithm achieved over 90% classification accuracy, a 10% improvement over non-TL methods.
• The proposed model outperformed traditional classifiers with minimal calibration data (two trials per gesture).
• High transfer robustness was demonstrated across different gestures (87.94% accuracy) and days (84.26% accuracy).

Conclusions:

• The proposed CNN TL method offers a practical solution to simplify the calibration process for sEMG-based HMI systems.
• This approach significantly improves gesture classification accuracy and reduces the burden on new users.
• Unsupervised TL shows great potential for advancing the commercial viability of sEMG-based HMI.