Design and Characterization of an Energy-Aware Embedded System for On-Device Postural Muscle Profiling
Abstract
Continuous postural muscle-stiffness monitoring can help prevent back and neck pain from prolonged sitting. Future wearables should process data on-device to reduce transmission energy and protect privacy, yet microcontroller (MCU)-class platforms must support efficient signal processing and machine learning under tight energy budgets. We design an energy-aware embedded system that performs Continuous Wavelet Transform (CWT), feature extraction, and Light-GBM inference on one MCU. We jointly study window length, CWT scale, tree count, and tree depth. On state-of-the-art vibration-sensing hardware and an ESP32-S3 Mini, a Paretooptimal configuration achieves 82.8% accuracy at 30 mJ per inference. Compared with BLE cloud offloading, it uses up to 7.2× less energy and requires no network connection.