Flexible Sensors for Biodynamic Studies: Enabling the In-Lab to On-Site Translation

Abstract

Conventional human motion or biodynamic studies often employ optical motion capture systems which are expensive and largely restricted to laboratory settings. Wearable sensors can provide a cost-effective, mobile alternative, particularly when miniaturized and integrated with a flexible design that allows direct, body-conforming adhesion to the skin. However, the potential of such flexible wearable sensors to advance naturalistic biomechanics and ergonomics research has not been explored nor assessed. This dissertation responds to this unmet need by a progressive series of three studies, (1) developing an approach for ambulatory gait measurement, (2) establishing a mobile biodynamic analysis system for human-exoskeleton interaction (HEI), and (3) exploring a deep learning-enhanced system for real-time lumbar spine joint load prediction. In the first study, an auto-calibrated algorithm based on principal component analysis was developed to measure human knee flexion/extension during level walking; in the second one, a flexible sensor-based system was built to capture lower body and lumbar spine joint kinematics and kinetics, and back muscle activity, during HEI. The algorithm and system were compared with conventional marker-based approaches, and the results endorsed that both approaches for measuring gait kinematics and for estimating lower body and lumbar spine joint kinematics and kinetics yielded comparable measurements to the marker-based method. The results also indicated that the proposed system for HEI was able to discern the effects of a passive low-back exoskeleton. Finally, predictive models to estimate lower body joint load during patient handling tasks using signals from flexible sensors and force plates were trained and tested with a customized convolutional neural network. An exhaustive search was performed to evaluate and compare all sensor combinations and practical solutions using three or four sensors to obtain accurate lumbar spine moment prediction were suggested. Collectively, this dissertation research establishes the feasibility of using flexible sensors to capture human motion, demonstrates the potential of deep learning to enable and enhance real-time minimum-input joint load prediction, and thus lays a foundation for flexible sensor-based field biomechanical studies.