An sEMG-Based Human-Exoskeleton Interface Fusing Convolutional Neural Networks With Hand-Crafted Features.

Area of Science:

• Rehabilitation Engineering
• Biomedical Signal Processing
• Human-Robot Interaction

Background:

• Surface electromyography (sEMG)-based human-robot interfaces (HRIs) are crucial for lower-limb exoskeleton robots in hemiplegic rehabilitation.
• Accurate lower-limb movement prediction for hemiplegic patients is challenging due to complex movement data and individual variability.
• Traditional methods using hand-crafted features offer limited insight, while deep learning methods may struggle with generalization.

Purpose of the Study:

• To develop an advanced human-exoskeleton interface for precise lower-limb movement prediction in hemiplegic patients.
• To overcome the limitations of existing methods by integrating diverse feature extraction techniques.
• To enhance the performance and generalizability of movement prediction in rehabilitation robotics.

Main Methods:

• Proposed a novel framework, HCSNet, fusing convolutional neural networks (CNNs) with hand-crafted features.
• Integrated time and frequency domain hand-crafted features with channel synergy learning-based features.
• Conducted sEMG data acquisition experiments to validate the HCSNet framework.

Main Results:

• HCSNet achieved high prediction accuracy: 95.93% in within-subject and 90.37% in cross-subject scenarios.
• Demonstrated superior prediction performance compared to existing lower-limb movement prediction methods.
• Validated the effectiveness of the fused feature approach in capturing complex movement dynamics.

Conclusions:

• The proposed HCSNet framework significantly improves lower-limb movement prediction accuracy for hemiplegic patients.
• Fusing CNNs with diverse hand-crafted and synergy-based features enhances feature representation and model generalization.
• This approach offers a promising advancement for intelligent lower-limb exoskeleton robots in rehabilitation.