Machine Learning on Microcontrollers

Description

Machine Learning is a very useful tool to analyze sensor data. It lets us automatically extract high-level information from data streams which may be complex, noisy and only weakly related to what we are interested in. With a bit of care around the computational requirements of the system, we can run the ML model directly on the sensor node. This lets us build affordable, power-efficient, privacy-compatible solutions which can operate stand-alone or be connected to other systems (IoT or otherwise).

My work in this area is centered on the open-source machine learning library emlearn. It consists of the emlearn C library (https://github.com/emlearn/emlearn) and the Micropython library emlearn-micropython (https://github.com/emlearn/emlearn-micropython).

I also collect relevant resources at: https://github.com/jonnor/embeddedml

Project Logs

Collapse

Toothbrush timer using accelerometer and ML
Jon Nordby • 01/04/2025 at 16:18 • 0 comments

I recently made an automatic toothbrush timer, using an accelerometer and a microcontroller. It is intended to be a fun and practical example for the emlearn-micropython machine learning library.

Below is a quick demo of the gadget in operation:

The project is available at: https://github.com/jonnor/toothbrush/
You will find there the software, firmware, data, ML training pipeline and 3d-printed parts.
Human Activity Recognition using MicroPython and emlearn
Jon Nordby • 01/04/2025 at 14:07 • 0 comments

There is now an example project in emlearn-micropython for doing (human) activity recognition using accelerometer/IMU. It can be found here: https://github.com/emlearn/emlearn-micropython/tree/master/examples/har_trees
Example applications of human activity detection could be sleep tracking, overall everyday activities' classification (sitting, walking, running, etc), exercise tracking (can be specific exercises like jacks/lunges/pushups/etc). Since machine learning is dependent on datasets, what exactly we will be able to do is of course a bit dependent on what datasets we are able to find and/or gather.

The feature extraction code is in MicroPython, but it could easily be ported to C. The classifier used is Random Forest, and that is implemented in C using the emlearn C library.
emlearn-micropython - Machine Learning and DSP for MicroPython
Jon Nordby • 01/04/2025 at 13:17 • 0 comments

emlearn-micropython is a Machine Learning and Digital Signal Processing library for MicroPython. It enables MicroPython applications to run efficient Machine Learning models on microcontroller,
without having to touch any C code. It builds on the emlearn C library and exposes a set of small, independent MicroPython modules. The modules (as .mpy files) can be installed at runtime, using the MicroPython package manager "mip". This aims to enable a user experience that is just as convenient and powerful as we already have on standard (C)Python, for microcontrollers.

Here is the first presentation on this library.
emlearn - a scikit-learn for microcontrollers
Jon Nordby • 01/04/2025 at 11:53 • 0 comments

emlearn is an open-source library for implementing. It allows to convert models trained with Python (using scikit-learn or Keras) into portable C99 code.
Here are two presentations that introduce the project. The first is very short and concise, where-as the other is a bit longer and detailed.