A Transfer Learning Based Cross-Subject Generic Model for Continuous Estimation of Finger Joint Angles From a New...

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Surface electromyography (sEMG) is crucial for human-machine interfaces (HMI), but subject-specific models lack generalizability.
• Inter-subject variability significantly degrades the performance of existing deep learning models for finger joint angle estimation.

Purpose of the Study:

• To develop a novel cross-subject generic (CSG) model for estimating continuous finger joint kinematics in new users.
• To improve the adaptability and performance of sEMG-based HMI systems across different individuals.

Main Methods:

• A multi-subject LSTA-Conv network was trained on sEMG and finger joint angle data from multiple subjects.
• A subjects adversarial knowledge (SAK) transfer learning strategy was employed to calibrate the model for new users.
• The CSG model's performance was validated on public Ninapro datasets.

Main Results:

• The CSG model demonstrated superior performance compared to subject-specific and other transfer learning models.
• Both the long short-term feature aggregation (LSTA) module and SAK transfer learning significantly contributed to the model's effectiveness.
• Increased diversity in the training set enhanced the model's generalization capabilities.

Conclusions:

• The proposed CSG model offers a robust solution for real-time finger joint angle estimation in HMI applications.
• This advancement holds significant potential for applications such as robotic hand control.
• The study highlights the importance of cross-subject learning strategies for personalized HMI development.