Deep learning for motion classification in ankle exoskeletons using surface EMG and IMU signals

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Ankle exoskeletons offer potential for mobility enhancement, rehabilitation, and fall risk reduction, especially in aging populations.
• Effective exoskeleton control relies on accurate, real-time prediction of user intentions from wearable sensor data.
• Existing sensor technologies can face challenges with comfort, durability, and usability for long-term deployment.

Purpose of the Study:

• To develop and evaluate a motion classification framework for ankle exoskeleton control.
• To integrate novel textile-based surface electromyography (sEMG) sensors with Inertial Measurement Units (IMUs).
• To assess the system's accuracy, adaptability, and robustness for real-world exoskeleton applications.

Main Methods:

• A framework combining three IMUs and eight sEMG sensors fabricated as comfortable, durable towel-based textile electrodes was developed.
• A dataset of multichannel time-series recordings for five functional daily motions was collected.
• Convolutional Neural Networks (CNNs) were employed for motion classification using the sensor data.

Main Results:

• The CNN-based framework achieved a classification accuracy of 99.263 ± 0.26%, significantly outperforming previous studies.
• Transfer learning demonstrated reliable adaptation to new users with minimal calibration data (ten samples per motion).
• The system maintained stable classification performance even when individual sensors were disrupted, indicating robustness.

Conclusions:

• The developed framework enables safe, high-accuracy, and real-world-ready ankle exoskeleton control.
• Deep learning combined with wearable textile electrodes and IMUs presents a feasible approach for advanced exoskeleton systems.
• The findings support the potential of this technology for improving user mobility and safety in daily activities.