Physics-informed hierarchical transformer for wearable sensor-based gait fatigue assessment

Gait-based fatigue assessment is important for sports injury prevention and rehabilitation monitoring, yet existing methods face limitations in accuracy and physical plausibility. Traditional approaches rely on handcrafted features that fail to capture complex spatiotemporal dependencies, while recent deep learning methods often produce predictions violating biomechanical principles.

This work presents a framework that integrates differentiable biomechanical constraints into hierarchical attention architecture for wearable inertial measurement unit (IMU)-based fatigue assessment. The method incorporates three components: (1) hierarchical multi-sensor attention that adaptively processes distributed IMU measurements through cross-sensor and temporal attention mechanisms; (2) differentiable biomechanical constraints implementing kinematic range limits, Newton-Euler dynamics, bilateral symmetry relationships, and mechanical energy conservation as learnable regularizers; (3) adaptive constraint weighting via curriculum learning that schedules physics enforcement from data-driven warmup to progressive constraint strengthening with fatigue-dependent scaling.

Evaluation on gait cycles from multiple participants demonstrates improved classification accuracy on multi-level fatigue assessment with robust performance under sensor noise and individual sensor failures. Cross-subject and cross-environment validation confirms generalization capability for field deployment.

This work advances the integration of physics-based reasoning with data-driven learning for biomechanical assessment in sports and rehabilitation applications.