Efficient and Dynamically Consistent Joint Torque Estimation for Wearable Neurotechnology via Knowledge Distillation.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning

Background:

• Wearable neurotechnology requires continuous movement monitoring for assessing motor function.
• Estimating joint torque from inertial measurement units (IMUs) on-device is challenging due to computational and energy constraints.
• Standard lightweight pipelines often omit complex signal processing, limiting accuracy.

Purpose of the Study:

• To develop an efficient on-device method for estimating joint torque from IMU data.
• To enable real-time movement analytics for wearable neurotechnology.
• To overcome limitations of existing lightweight pipelines in capturing complex dynamics.

Main Methods:

• Proposed a Physically Guided Dual-Consistency Knowledge Distillation (PDC-KD) framework.
• Integrated biomechanical priors via parameter-manifold alignment and physics-guided compensation.
• Utilized Fisher-information-weighted parameter transfer and a physics-guided regularization term for dynamic consistency.

Main Results:

• The student model achieved teacher-level predictive accuracy with significant resource reduction (98% parameter reduction, ~2x faster inference, ~1 ms latency).
• The method demonstrated enhanced dynamical consistency in torque estimates.
• The framework operates effectively within the resource constraints of edge devices.

Conclusions:

• PDC-KD provides an efficient solution for on-device joint torque estimation in wearable neurotechnology.
• The approach enables reliable real-time movement analytics for improved motor impairment characterization and recovery monitoring.
• This method addresses the critical need for accurate kinetic marker estimation in resource-limited edge computing environments.