TinyML is the set of techniques that make it possible to run machine learning models on microcontrollers with just a few hundred kilobytes of memory. Here is the path, with no shortcuts.

1. Start from a small, measurable problem

"Understand whether the pump works well" is too vague. "Tell normal operation, cavitation and a worn bearing apart from vibrations" is a problem a small model can solve. Define clear classes and a success criterion, for example fewer than 2% false alarms.

2. Collect data with the final sensor

The model learns the characteristics of the sensor it was trained with. Use the same board, the same mounting and the same conditions you will have in production, and also collect the "boring" cases: the machine switched off, background noise, ambiguous situations.

3. Extract the right features

For audio and vibrations you almost always work on frequency-domain representations, which make the problem easier for a small network. Good data preparation is worth more than a bigger network.

4. Train, quantise, measure

After training, the model is quantised to 8 bits. Always check accuracy after quantisation, the memory used and the inference time on the board: those are the numbers that matter.

5. Test in the field

Leave the prototype installed for days. Record the cases where it is wrong and add them to the dataset: that is how a good prototype becomes a reliable product.