On-Device Energy Tracking from Phone Accelerometer and Thermal Sensors
A lightweight ML model runs entirely on-device, estimating personal energy levels throughout the day without cloud uploads.
By analyzing movement patterns, sleep quality, and device thermal signatures, a new on-device app predicts energy dips and peaks, helping users identify when they're most productive or need a break.
The prototype, developed during a hackathon, uses a quantized lightweight neural network to process accelerometer and thermal sensor data in real time. The model is trained on self-reported energy scores correlated with movement patterns, achieving a mean absolute error of 3.2 on a 0–100 scale.
In a day-long demo, participants wore the app while working on coding tasks. The predicted energy scores closely tracked self-reported levels: by mid-morning, the index climbed to , up from at 9 AM. After lunch, a dip to was observed, followed by a recovery to in the late afternoon.
The app's timeline dashboard shows a simple line chart of energy predictions over the day, with annotations for sleep periods and thermal spikes. No data leaves the device—all processing happens on-device using Apple's Core ML and Android's NNAPI.
The team behind the project, participating in the Edge AI Hackathon, plans to open-source the model and dataset. They emphasize that the technology is not a medical device but a productivity tool for self-awareness.