On this page
The Silent-Drift Problem: False Confidence Until Calibration Day
A monitoring probe rarely fails loudly. It drifts. Each month it reads a little further from the truth, and because the displayed number still sits comfortably inside your 2 to 8C band, nobody questions it. The cold room looks fine. The daily logs look fine. The team signs off shift after shift with full confidence. Then calibration day arrives, the calibration house reports the probe is reading 0.7C high and out of tolerance, and you realise the confidence was false. The readings were biased for months, and every stored-product record they support is now in question.
That is the trap of slow drift. A sudden sensor failure trips an alarm and gets investigated the same day. A gradual bias produces clean-looking logs right up to the moment it is caught, and by then the damage is retrospective: you cannot un-store the product that sat under a probe you now know was wrong. PPB Good Distribution Practice expects monitoring equipment to be calibrated at defined intervals and held within stated tolerance, and an inspector reading an out-of-tolerance calibration certificate will ask the obvious question: how long was it drifting, and what did that do to the records in between?
The Sensor Drift & Calibration Alert Tool answers that question before calibration day forces it. It compares a probe against a trusted reference over time, measures the bias, measures how fast that bias is growing, and projects the date the probe will cross your acceptable tolerance. If that date lands before your scheduled calibration, it tells you to pull the calibration forward. This guide walks one real-shaped example end to end and shows how to write the result up so it survives an audit. It is a companion to your calibration and monitoring SOPs, not a replacement for them.
Worked Example: A Probe Drifting 0.6C High
Take a distributor running a 2 to 8C cold room in Nairobi. The room has its routine monitoring probe and, alongside it, a reference sensor used for periodic verification checks. The acceptable tolerance for the monitoring probe against the reference is plus or minus 0.5C. The probe was last calibrated 120 days ago, and its next scheduled calibration is 60 days out.
Pull three months of paired readings (the monitoring probe and the reference, on aligned timestamps) and the pattern is not flat. The bias is real and it is growing:
- Month 1: the probe reads on average about 0.2C above the reference. Easy to dismiss as noise.
- Month 2: the average bias is about 0.4C high. Still inside tolerance, still invisible on the daily logs.
- Month 3: the average bias is about 0.6C high. The probe has now quietly crossed the plus or minus 0.5C tolerance, and nobody on the floor can see it because the displayed temperature still reads inside 2 to 8C.
The tool detects two things at once: a current mean bias of about plus 0.6C, and an upward drift trend of roughly 0.2C per month. It does not stop at "the probe is biased". It fits the trend and projects forward. The bias was crossing the plus or minus 0.5C line near the end of month 3, and on the current slope it keeps climbing. Where calibration discipline matters is the comparison against your schedule: calibration is booked for 60 days from now, but the projection shows the probe will be meaningfully out of tolerance in roughly 35 days, well before that date.
So the recommendation is concrete: pull the calibration forward, into the next two to three weeks, rather than waiting the full 60 days. Acting on that recommendation does two things. It stops the probe accumulating another month of biased readings that would later need an out-of-tolerance impact assessment, and it gives you a defensible record showing you detected the drift from data and responded before tolerance was breached, instead of discovering it on calibration day. You did not wait for the probe to invalidate a quarter of stored-product records. You caught the trend and moved.
What Data You Need and Why
Four inputs drive the analysis. The core pair is the probe readings and a trusted reference on aligned timestamps. The reference can be a calibrated reference sensor, a second independent probe in the same space, or periodic verification check readings taken against a calibrated thermometer. Without something to compare against, the tool can only show what the probe says, not whether what it says is true. The whole method rests on the gap between the two.
The second input is time span. Drift is a trend, and a trend needs length to be visible. Two or three months of paired readings is the practical minimum; a single week of data can show a bias but cannot reliably show that the bias is growing. The longer and more regular the comparison history, the more confident the projected breach date.
The third input is your tolerance spec: the acceptable deviation of the probe against the reference, taken from your calibration SOP, for example plus or minus 0.5C. This is the line the projection is aimed at. The fourth is the last calibration date (and the next scheduled date), so the tool can place the projected breach against your existing schedule and tell you whether to wait or pull it forward.
The minimum file is timestamp, probe reading, and reference reading, headers on row 1, one paired reading per row, consistent units throughout. Export raw interval readings from your logger platform rather than chart screenshots or daily averages; averaging smooths the very signal you are trying to measure. If the export is messy or the two sensors are on slightly different timebases, clean and align it through the Data Logger QA Cleaner before uploading here.
Reading the Outputs
The tool returns four results. Read them in order:
- Current bias: the probe's present average offset from the reference, for example plus 0.6C. This is where the probe stands today. On its own it tells you whether you are already inside or outside tolerance, but not where you are heading.
- Drift rate: how fast the bias is changing, for example 0.2C per month. This is the output that separates a stable but slightly offset probe (live with it, recalibrate on schedule) from one that is actively going wrong (act now). A near-zero drift rate with a small bias is far less urgent than a small bias climbing steadily.
- Projected tolerance-breach date: the trend carried forward to the day the bias is expected to cross your tolerance spec. This is the number you compare against your calendar. A breach date inside your current calibration interval is the trigger to act.
- Recalibration recommendation: the plain-language call. If the projected breach lands before your scheduled calibration, the tool recommends pulling it forward and gives you the rough window. If the probe is stable and well inside tolerance, it confirms the existing schedule is adequate.
Never read these in isolation. Pair them with operational context: a step change in bias the week after a maintenance visit points at a knocked or repositioned probe rather than gradual ageing, and a bias that only appears during defrost cycles is a placement question, not a calibration one. If the offset looks more like a hotspot than a true drift, a cold room mapping study will show whether the probe is simply sitting in a colder or warmer pocket of the room. And once you have the recalibration call, your alarm thresholds should be reviewed too, because a biased probe may have been triggering, or suppressing, alarms against the wrong baseline.
Documenting the Change for PPB and GDP Audits
Pulling a calibration forward is a change to a scheduled control, and a probe found drifting has implications for the records it produced. Capture both the way an inspector expects to see them:
- Attach the raw paired export (probe and reference) and the cleaned file used for the analysis, with the comparison window dates.
- Record the measured current bias, the drift rate, the tolerance spec applied, and the projected breach date, alongside the last calibration date and the original schedule.
- Raise an out-of-tolerance impact assessment if the projection shows the probe was already across tolerance during any period when product was stored, and assess what those biased readings mean for that product. The excursion impact assessor helps quantify the effect of the corrected temperatures.
- Route the schedule change through change control with QA sign-off, stating that the decision to pull calibration forward was driven by a documented drift trend, not convenience.
- Re-review at the next management review using the actual calibration result, and confirm whether the prediction matched what the calibration house found.
That chain (data, projection, decision, calibration outcome) is what makes the action defensible. It turns "the probe was out of tolerance on calibration day" into "we detected the drift from monitoring data and recalibrated before it breached tolerance", which is exactly the proactive control GDP rewards. If you are still building these records, our GMP readiness checklist covers the calibration, deviation, and change-control documentation an inspector will ask for.
Frequently Asked Questions
How is this different from just waiting for the next scheduled calibration?
Scheduled calibration tells you the probe was out of tolerance after the fact, once months of biased readings already exist. This tool projects the breach date from the drift trend in your own data, so you can recalibrate before tolerance is crossed instead of discovering the problem on calibration day and then having to assess every record in between.
What can I use as the reference if I do not have a calibrated reference sensor?
A second independent probe in the same space, or periodic verification check readings taken against a calibrated thermometer, both work. The method only needs a trusted comparison point on aligned timestamps. The more reliable the reference, the more trustworthy the projected breach date.
How much history do I need before the projection is reliable?
Drift is a trend, so it needs length. Two to three months of paired readings is the practical minimum. A single week can reveal a current bias but cannot reliably show that the bias is growing, which is the part that drives the projected breach date.
The probe is biased but stable, with almost no drift. Do I still act now?
Not necessarily. A small, steady offset that stays inside tolerance can usually wait for the scheduled calibration. The urgent case is a bias that is climbing: a low drift rate with a comfortable margin is very different from a small bias growing every month toward the tolerance line. Read the drift rate alongside the current bias.
If the projection shows the probe was already out of tolerance, what do I do about past records?
Raise an out-of-tolerance impact assessment for the period the probe was across tolerance while product was stored, and evaluate what the corrected temperatures mean for that product. Document the decision through change control. This guide supports the analysis, but final disposition must follow your approved SOPs and QA sign-off.
Check Your Own Probes for Drift
Upload two to three months of paired probe and reference readings, set your tolerance, and get a current bias, drift rate, projected tolerance-breach date, and recalibration recommendation you can take straight into change control.
Open Sensor Drift Tool