Towards Integration of Domain Knowledge-Guided Feature Engineering and Deep Feature Learning in Surface...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Surface electromyography (sEMG) signal recognition is crucial for controlling noninvasive neural interfaces like prosthetic limbs and rehabilitation robots.
• Current deep learning methods for sEMG recognition face limitations due to signal noise and non-stationarity, despite advancements in feature learning.
• Feature engineering has been explored to mitigate these limitations, but achieving a balance between performance and computational cost remains a challenge.

Purpose of the Study:

• To develop a novel framework, the progressive fusion network (PFNet), for enhanced sEMG-based hand movement recognition.
• To improve recognition accuracy while managing computational complexity by integrating domain knowledge with deep feature learning.
• To provide a robust solution for controlling myoelectric prostheses and rehabilitation robots.

Main Methods:

• Proposed a progressive fusion network (PFNet) framework that synergistically combines domain knowledge-guided feature engineering and deep feature learning.
• Utilized a feature learning network for high-level representations from raw sEMG signals and a domain knowledge network for engineered time-frequency features.
• Implemented a 3-stage progressive fusion strategy to integrate features from both networks for final decision-making.

Main Results:

• PFNet achieved average hand movement recognition accuracies of 87.8%, 85.4%, 68.3%, 71.7%, and 90.3% across five diverse sEMG datasets.
• The proposed PFNet significantly outperformed existing state-of-the-art methods in sEMG-based hand movement recognition.
• Demonstrated the effectiveness of integrating engineered features with deep learning for improved performance and potential for computational efficiency.

Conclusions:

• The progressive fusion network (PFNet) offers a superior approach to sEMG-based hand movement recognition.
• This hybrid method effectively addresses the challenges posed by noisy and non-stationary sEMG signals.
• PFNet shows significant promise for advancing the control capabilities of noninvasive neural interfaces.