A self-powered and self-sensing multi-joint wearable energy harvester with lightweight edge-deployable CNN for real-time lower-limb kinematics and motion mode recognition

Abstract

With the global aging population, continuous monitoring of daily activities is increasingly important for rehabilitation and health assessment. However, conventional wearable devices are plagued by insufficient energy self-sufficiency, limiting their practical application in long-term activity monitoring. In this work, we propose a self-powered and self-sensing multi-joint wearable energy harvester capable of simultaneously harvesting negative mechanical work from both the hip and knee joints during human locomotion. The harvested biomechanical energy is converted into electrical power (0.39 W per device) to support embedded computing and wireless communication modules, enabling intermittent energy-autonomous operation. Meanwhile, the generated electrical signals are directly utilized for motion sensing without additional sensors. To achieve real-time motion analysis on resource-constrained embedded platforms, a lightweight convolutional neural network optimized via multi-objective optimization is developed for edge deployment. The proposed framework simultaneously estimates hip and knee angular velocities and recognizes seven motion modes. Experiments involving five participants demonstrated accurate lower-limb kinematics estimation with a mean absolute error of 0.05 ± 0.01, root mean square error of 0.09 ± 0.02, and correlation coefficient of 0.96 ± 0.01, while achieving a motion recognition accuracy of 98.7%. This work establishes a new paradigm integrating biomechanical energy harvesting, self-sensing, and embedded artificial intelligence for sustainable wearable rehabilitation and daily activity monitoring.