How Often to Recalibrate, and How to Know
Annual calibration is a convention, not an answer. The interval should come from observed drift and from what a wrong reading costs — and most plants have the data to work it out.
Most plants calibrate instruments annually. Almost none can say why annually. The interval was inherited, or it matches the audit cycle, or a vendor suggested it, and it is applied uniformly to a pressure transmitter on a critical safety loop and a temperature gauge on a storage tank nobody reads. That uniformity is the problem: it over-calibrates most instruments, under-calibrates a few, and spends the budget in the wrong places.
Start with what calibration actually establishes. It is a comparison of the instrument against a reference of known and better accuracy, producing a statement about error at several points across the range. Two results matter and they are routinely conflated. The as-found data — what the instrument read before anything was adjusted — is the valuable one, because it tells you what the instrument has been doing since the last calibration and therefore whether every measurement in that period was trustworthy. The as-left data merely confirms the technician did their job. A calibration programme that records only as-left has thrown away the information it was collecting.
That as-found history is what sets the interval. Accumulate it and you can see the actual drift behaviour of that instrument type in that service: how far it moves in a year, whether the movement is a steady trend or a random walk, whether it correlates with process upsets or with cleaning. Then the interval follows from a policy — for example, calibrate at an interval over which the instrument is expected to remain within a stated fraction of its tolerance — rather than from a habit. Instruments that never move get longer intervals, and instruments that do get shorter ones. This is the substance of the various interval-adjustment schemes in the metrology literature, and it does not need any of them to be useful: the simple version, lengthen the interval when several consecutive as-found results are well within tolerance and shorten it when one is not, captures most of the benefit.
The second input is consequence, which is where the interval stops being a purely technical question. A wrong reading on a custody transfer meter is money leaving the site continuously. A wrong reading on a safety instrumented function is a protection layer that is not there, and the proof test interval is usually set by the SIL calculation rather than by drift at all. A wrong reading on a regulatory emissions monitor is a compliance exposure. A wrong reading on a tank level that an operator sanity-checks against a sight glass daily is a nuisance. These four instruments should not share a calibration interval, and ranking instruments by consequence is usually a morning's work that permanently reallocates the maintenance budget.
Several practical points decide whether any of this survives contact with the plant. Calibrate in place where the installation is part of the measurement — a thermowell's response, an impulse line that is partly plugged, an orifice plate installed backwards, a flow meter with insufficient straight run — because a transmitter that is perfect on the bench and wrong in the pipe is a very common outcome, and bench calibration will never find it. Check the reference: the standard used must be traceable and must itself be in calibration, with an accuracy ratio comfortably better than the instrument under test, and an expired reference invalidates every calibration performed with it, which is a discovery nobody wants to make during an audit.
Record uncertainty, not just a pass or fail. A pass means the error was within tolerance according to a measurement that itself has error, and where the result sits close to the limit the honest statement includes the uncertainty. This matters most in exactly the cases where the calibration is being used to make a decision.
And watch for drift that is not drift. An instrument that suddenly reads differently has often not drifted; it has been affected by something physical — a plugged line, a coating on a sensor, a leaking seal, a change in ambient temperature at the transmitter, a wiring fault. Adjusting it back to the reference makes the number look right and leaves the fault installed. The as-found record is what distinguishes the two: a slow, consistent trend is drift, and a step change is an event.
Finally, treat the schedule as a live document with a small feedback loop: record as-found, compare to tolerance, adjust the interval, review the outliers. A programme that does that will typically extend intervals on most of its instrument population, shorten them on a handful, and find two or three instruments that have been quietly wrong for a year — which is the entire justification for doing it at all.