A Deep Transfer Learning Approach to Reducing the Effect of Electrode Shift in EMG Pattern Recognition-Based Control
Summary
This study introduces transfer learning (TL) with convolutional neural networks (CNNs) to improve prosthetic control. The method enhances robustness against factors like electrode shift, enabling faster recalibration for better electromyogram (EMG) pattern recognition.
Area of Science:
- Biomedical Engineering
- Machine Learning
- Rehabilitation Technology
Background:
- Commercialization of myoelectric prostheses is hindered by pattern recognition system fragility.
- Factors like electrode shift and skin impedance variations degrade performance.
Purpose of the Study:
- To develop a robust supervised adaptation method for electromyogram (EMG) pattern recognition.
- To improve the commercial viability of myoelectric prostheses through enhanced adaptability.
Main Methods:
- Proposed a transfer learning (TL) approach using convolutional neural networks (CNNs).
- Validated the method with able-bodied subjects experiencing simulated electrode shifts.
- Fine-tuned pre-trained CNNs with minimal post-shift calibration data.
Main Results:
- The CNN TL method significantly outperformed training from scratch and support vector machines (SVMs) with limited data.
- Achieved superior performance compared to previous linear and quadratic discriminant analysis (LDA/QDA) adaptation techniques.
- Demonstrated effective recalibration within seconds for each class.
Conclusions:
- The proposed CNN TL approach offers a practical solution for adapting EMG pattern recognition systems to external perturbations.
- Enhances the robustness and reliability of myoelectric control systems.
- Facilitates the commercialization of advanced prosthetic technologies.
