← all projects

project

Battery vibration node that predicts bearing wear

An ESP32 sensor running a quantised classifier on-device, so a two-year battery life survives contact with real motors.

Placeholder cover — swap this for a photo of the sensor mounted on a motor housing.
Date
Nov 2025
Role
Firmware and model deployment
Status
shipped
Built with
ESP32-S3, C++, ESP-IDF, TensorFlow Lite Micro, MEMS accelerometer

Streaming raw accelerometer data off a battery-powered node is a good way to have a dead node by week three. The radio costs far more than the maths does, so the classification has to happen where the data already is.

Approach

Sample a window, run a small quantised model on-device, transmit a verdict and a confidence figure. The full waveform is only sent when the model is uncertain, which turns out to be rarely.

Result

Average current draw landed at 61 µA, giving just over two years on a pair of AAs. The model catches early-stage outer-race wear about three weeks before the vibration is obvious to a technician with a handheld meter.

Edge AIFirmwarePredictive Maintenance