Estimation of Joint Angle From sEMG and Inertial Measurements Based on Deep Learning Approach.

Area of Science:

• Biomedical Engineering
• Robotics
• Human-Machine Interaction

Background:

• Continuous kinematics estimation from surface electromyography (sEMG) is crucial for intuitive human-machine collaboration.
• Multimodal inputs, including sEMG and inertial measurements, show promise for enhancing estimation performance.

Purpose of the Study:

• To assess the efficacy of combining angular velocity and myoelectric signals for continuous prediction of 12 hand joint angles.
• To evaluate estimation performance across five functional and grasping movements in 20 subjects.

Main Methods:

• Utilized convolutional and recurrent neural networks with transfer learning (TL).
• Employed a pretrained deep network model, adapted from basic hand movements to functional motions.
• Compared multimodal input (sEMG + angular velocity) against sEMG-only methods.

Main Results:

• The multimodal approach showed a slight performance improvement over sEMG-only methods for most movements.
• Both strategies exhibited similar estimation behavior across various tasks.
• A significant improvement was observed specifically for the bottle-opening task using a tripod grasp.

Conclusions:

• Combining angular velocity with sEMG offers marginal benefits for continuous hand kinematics estimation.
• Transfer learning with deep neural networks is effective for adapting models to functional hand movements.
• Specific functional tasks, like tripod grasp bottle opening, can benefit substantially from multimodal sensing.