Context
Measuring instruments are calibrated on a fixed schedule, and between visits their condition is unknown. That leaves two situations, and most sites live with both.
An instrument is pulled and calibrated while still comfortably within tolerance; the calibration still produces the record the quality system requires, but the interval could have been longer without added risk.
Or an instrument is found out of tolerance, and because there is no way to tell when it crossed the limit, the organisation has to judge whether measurements taken since the previous calibration were affected, and act on that judgement.
The schedule is set from convention rather than from the condition of the individual instrument.
What we tested
At each calibration the as-found error is recorded against a known reference. That sequence of recorded errors is the input.
The model estimates how much service life remains before the instrument reaches its tolerance limit, and it reports a prediction window rather than a single date. It is built on a class of lightweight machine-learning techniques (reservoir computing) suited to time-series signals and edge deployment. Validated on a publicly available multi-year record of instrument drift.
What it shows
Once drift has begun, the remaining service life can be estimated well ahead of the tolerance limit, with a stated prediction window.
The estimate is also traceable: the model shows which part of the recent calibration history is driving it.
This may apply to your instruments
If you calibrate instruments on a fixed schedule and record the as-found error each time, that record already contains what this approach needs.
It can support moving an instrument's calibration earlier when its condition warrants it, and later when it does not.
A short conversation is usually enough to tell whether your calibration records are in a usable form.