Improved Transfer Learning for Detecting Upper-Limb Movement Intention Using Mechanical Sensors in an Exoskeletal...

Area of Science:

• Robotics
• Biomedical Engineering
• Machine Learning

Background:

• Detecting upper-limb motion intentions is crucial for human-robot collaboration, especially in rehabilitation.
• Traditional methods often rely on complex sensor arrays, including surface electromyography (sEMG).
• There is a need for simpler, yet accurate, sensor-based intention detection systems.

Purpose of the Study:

• To propose a novel strategy for detecting upper-limb motion intentions.
• To utilize deep and heterogeneous transfer learning techniques with mechanical sensor signals.
• To evaluate the performance of models trained with reduced sensor inputs.

Main Methods:

• Combined surface electromyography (sEMG), force-sensitive resistors (FSRs), and inertial measurement units (IMUs) for signal capture.
• Developed deep learning models (CIFAR-ResNet18, CIFAR-MobileNetV2) for intention detection.
• Employed transfer learning, training a target model using only FSR and IMU signals, with optimized layer structures and learning rates.

Main Results:

• The source model using CIFAR-ResNet18 achieved 95% accuracy and a 0.95 F-1 score.
• The optimized target model, using only FSR and IMU signals, reached 93% accuracy and a 0.93 F-1 score.
• Mechanical sensors alone demonstrated performance comparable to models incorporating sEMG.

Conclusions:

• Mechanical sensors (FSRs and IMUs) can effectively detect upper-limb motion intentions.
• Deep and heterogeneous transfer learning offers a precise and convenient approach for intention detection.
• The proposed algorithm is suitable for human-robot collaboration in rehabilitation assistant robots.