Vibration anomaly detection on a microcontroller

The brief

Sending raw vibration data to the cloud is slow and costly; a sensor that decides for itself is better. Turn vibration signals into features, train a small anomaly or fault model, shrink it to fit a microcontroller, and measure accuracy, latency, memory and power on the device. Good looks like: features computed on the device, a model within memory limits, an honest accuracy comparison with the laptop version, and a demo on a shaking test rig.

What you will build

1. Prepare windows of vibration data and compute features
2. Train a small classifier or autoencoder
3. Quantise and convert it for the microcontroller
4. Run it on the board and measure latency, memory and power
5. Compare accuracy on the device with the laptop
6. Show a live demo on a simple vibration rig

Free, open data to build it with