Why we built ML sensor calibration
The $3 sensor problem
A $3 AliExpress pH probe and a $100 commercial one measure the same voltage. What you pay for is the calibration, the temperature compensation, and the drift-correction — the software wrapped around the electrode. So we built that software and pointed it at the cheap probes. This is why Verdalio ships ML sensor calibration, and how it makes budget hardware behave.
What actually goes wrong with cheap probes
- Offset drift — the probe reads 6.4 when the solution is truly 6.0.
- Slope error — it tracks changes, but compresses or stretches the scale.
- Temperature sensitivity — EC in particular swings with temperature.
- Aging — every probe degrades; the error grows over weeks.
None of these make the sensor useless. They make it uncorrected. Correction is a learnable function.
The calibration model
The hub runs a small local model (a linear baseline that the ML layer only improves on when it clears a cross-validation gate) mapping raw probe readings to true values, using your two- or three-point calibrations plus ongoing drift signals. It runs on the Raspberry Pi via a localhost service — no cloud, no data leaving your network.
Why this matters for the whole platform
Cheap-sensors-smart-software is not a slogan, it is the economic model that lets a hobbyist build lab-grade automation for hobby money. It is why a beginner can start with a $10 probe and still trust the pH graph — and why Dávid can wire up a dozen DIY nodes without buying a dozen commercial instruments.
Learn how to calibrate your probes in the Devices & Sensors guide.