AI in quality

# Predictive maintenance in GMP: how CMMS, calibration and reliability data decide when equipment is serviced, calibrated and replaced
By QLogix · Published 15 October 2026 · 12 min read
The short answer
A preventive interval copied from a manual is a guess. The work orders, downtime and calibration results a site already records contain the answer: mean time between failures, availability, whether failures are rising with age, and the interval that catches most of them before they happen. When the CMMS, the calibration module and the quality system share one record graph, the model reads itself, the engineer confirms, and the change goes through change control.
Failures on one asset before a Weibull fit of the times between them says whether they are infant, random or wear-out. 4
The life by which 10 percent of units have failed: a preventive interval set there catches nine failures in ten for a High-criticality asset. B10
Signatures a GMP calibration needs: the second-person review of the readings, and QA's decision when a reading is out of tolerance. 2

## Key takeaways
- 21 CFR 211.67 and 211.68 require written maintenance and calibration programmes with records of what was done; 21 CFR 820.70(g) and EU GMP Chapter 3 ask the same of devices and of European sites. None of them says how often. The interval is the firm's decision, and the records the firm keeps are the evidence that it was the right one.
- The maintenance history already holds the reliability figures: corrective and emergency work orders are failures, their downtime is unavailability, their labour is the corrective share. Mean time between failures, mean time to repair, availability and the failure trend fall out of the records without a sensor on the machine.
- With four or more failures, a Weibull fit of the times between them gives the shape: below 1 the asset is failing young and a shorter interval will not help; near 1 failures are random and the plan fits; above 1.3 failures rise with age and the interval should shorten toward the B-life for the asset's criticality, or the replacement case should be written.
- Calibration belongs in the same model. An out-of-tolerance reading is a failure of the measurement, with its own impact assessment on what was measured since the last good calibration, a second-person review of the readings and QA's signed decision. On a linked equipment file, it counts toward the asset's reliability and its lifecycle stage.

## What do the regulations require of maintenance and calibration?
For drugs, 21 CFR 211.67 requires that equipment be cleaned, maintained and, where appropriate, sanitized at appropriate intervals, under written procedures, with records of maintenance, cleaning and inspection kept. Section 211.68 covers automatic, mechanical and electronic equipment: routinely calibrated, inspected or checked according to a written program designed to assure proper performance, with written records of those checks maintained. Section 211.160(b)(4) adds the calibration of instruments, apparatus, gauges and recording devices at suitable intervals, under a written program with specific directions, schedules, limits for accuracy and precision, and remedial action when limits are not met. For devices, 21 CFR 820.70(g) requires equipment to be maintained under schedules with the activities and dates documented, and calibration under 820.72 with the accuracy and precision limits and the remedial action stated.
EU GMP Chapter 3 asks that equipment be designed, located and maintained to suit its intended purpose, that repair and maintenance not present a hazard to product quality, and that measuring, weighing, recording and control equipment be calibrated and checked at defined intervals by appropriate methods with records kept. Annex 15 ties the qualified state of equipment to its maintenance and calibration: a change that affects it goes through change control and, where needed, requalification.
What the regulations do not say is how long the interval should be. 'Appropriate', 'suitable' and 'defined' leave the number to the firm, which makes the interval a decision that needs a rationale and the records that show the rationale held. That is where the maintenance history comes in: a firm that can show mean time between failures, availability and the trend for each asset has a rationale for its intervals, and a firm that cannot is relying on the manufacturer's leaflet.
What the regulations ask, and what the records must show
Expectation Where it comes from What the connected record shows
Maintenance at appropriate intervals under written procedures, with records 21 CFR 211.67; 820.70(g); EU GMP Chapter 3 A preventive plan per asset with its tasks, intervals and rationale; every work order with its tasks, labour, parts and verification
Calibration under a written program with limits for accuracy and precision and remedial action 21 CFR 211.68; 211.160(b)(4); 820.72 An approved calibration requirement per asset: interval and basis, points and tolerances, traceability, the out-of-tolerance rule
Records of the checks, with the result and the person 21 CFR 211.68(a); 211.194 As-found and as-left readings judged by the system, the certificate, the reviewer's signature
Remedial action when limits are not met 21 CFR 211.160(b)(4); 820.72(a) The impact assessment on what was measured since the last good calibration, QA's signed decision, the investigation linked
The qualified state maintained through change control EU GMP Annex 15 A plan change with its reason in the audit trail; a change control where the qualified state is affected
Data integrity of the records FDA data integrity guidance Attributable entries, a hash-chained audit trail, signatures with password re-entry

A reading of 21 CFR 211.67, 211.68 and 211.160(b)(4), 21 CFR 820.70 and 820.72, and EU GMP Chapter 3 and Annex 15. The wording of each expectation is ours; the sources are listed at the end.

## One equipment file: CMMS, calibration and the quality system on the same record graph
The reason predictive maintenance stays a slide at most sites is that the data lives in three places. The CMMS holds the work orders, the calibration system holds the certificates, and the quality system holds the deviations that the failures caused. Each knows one third of the asset's story, and the engineer who wants the whole of it exports three spreadsheets. By the time they are joined the quarter is over.
A connected architecture treats the maintenance asset as a node with everything about it linked both ways: its preventive plan and work orders, the calibration records of the instruments and gauges mounted on it, the deviations opened when it failed, the change controls that altered it, the validation project that qualified it. A link is an audited object with a direction and a meaning, not a column in a spreadsheet, so the inspector follows it as easily as the engineer. The figure shows the flows in QLogix: the maintenance asset at the centre, work orders feeding it failures and downtime, calibration records feeding it out-of-tolerance events, both feeding the quality system when a result or a batch is affected, and the reliability model and the AI reading across all of it under the user's own permissions.
The data integrity expectations apply across the graph as they do inside any one record. Each work order and each calibration is recorded by the person who did or received it, time-stamped by the system, judged by the system where there is a limit to judge against, reviewed by a second person where the regulation expects one, and kept with its history in a hash-chained audit trail. Nothing in the model below reads a figure that was not recorded that way.
The equipment file: maintenance, calibration, quality and the reliability model on one record graph
- data systems
- AI
- people decide and sign
- connected systems

The dots travel in the direction of the data. Work orders and calibration records feed the maintenance asset; out-of-tolerance results and failures open investigations in the quality system; change controls alter the plan; the reliability model and the AI read across the graph and write only drafts; every arrow is an audited link.

## Reading the maintenance history: MTBF, MTTR, availability and the trend
The first figures need no statistics. Every corrective or emergency work order completed on an asset is a failure; the downtime recorded on it is unavailability; the labour booked to it is the corrective share of the asset's maintenance. From the first record to today, the mean time between failures is the operating time divided by the number of failures, the mean time to repair is the mean downtime per failure, and the availability over the last year is one minus the downtime over the hours in the window. The failure trend compares the last 180 days with the 180 before: rising, steady or falling.
Operating time is the figure most sites get wrong. A filling line that runs one shift and a bioreactor that runs continuously should not be compared in calendar days. An asset's operating hours per day, or its run-hour counter read on a known date, convert the history into operating time, and the MTBF comes out in hours the equipment actually ran. Without it the model says so and reports calendar days, which overstate the life of equipment that sits idle.
These four figures are enough to see the fleet. The asset with an availability under 90 percent and a corrective share over 60 percent is already a replacement case, whatever its age. The asset whose recent failures are double the previous period's is wearing out. The asset with two failures in its first year, both within ninety days of installation, is in early life, and a shorter preventive interval will not fix an installation fault.
Failures per quarter on one asset, and the point at which the trend becomes visible

An illustrative count of corrective and emergency work orders per quarter on one asset over five years, not a measurement. The band is the rate the plan was written for; the run of rising quarters is what the trend rule and the Weibull shape pick up two quarters before the quarter that breaks the limit.

## When is an asset wearing out? The Weibull shape and the bathtub curve
Counting failures tells you how many; the times between them tell you why. The Weibull distribution describes the three regimes of the bathtub curve with one shape parameter, beta. Below 1, the failure rate is falling with age: the asset fails young, from installation, commissioning, parts quality or operator error, and gets more reliable as those are fixed. Near 1, failures are random, independent of age, and a preventive interval cannot reduce them; condition monitoring and spares can. Above 1, the failure rate rises with age: wear, fatigue, fouling, drift. That is the regime a preventive interval is for, and the further beta is above 1, the sharper the knee and the more a well-placed interval saves.
The fit needs data: with fewer than four failures the times between them cannot be ranked meaningfully, so the model holds its advice to the MTBF and the trend until there are enough. With four or more, median-rank regression on the log of the times gives beta and the characteristic life eta, the time by which 63 percent of units have failed. From those, the B-life for any probability follows: B10 is the time by which 10 percent have failed. A preventive task placed at B10 lets through one failure in ten; at B5, one in twenty.
The criticality sets which B-life the interval targets. In QLogix a Critical asset's preventive interval is set at B5, a High asset's at B10, a Medium asset's at B20 and a Low asset's at B30; without a Weibull fit, the same idea is a fraction of the MTBF. The lifecycle stage follows from beta and the trend, then from availability and the corrective share: early life, useful life, wear-out, or end of life when wear-out comes with lost availability or a maintenance load that is mostly corrective.
From the work orders to the lifecycle stage
- 1 Failures and downtime
Corrective and emergency work orders completed on the asset, and out-of-tolerance calibrations on its linked instruments; the downtime and labour recorded on them.

- 2 Operating time
Hours per day or the run-hour counter convert calendar days to operating hours; without them the model says it is working in calendar days.

- 3 MTBF, MTTR, availability, trend
Operating time per failure, mean downtime per failure, one minus downtime over the last year, the last 180 days against the 180 before.

- 4 Weibull fit
With four or more failures: beta (the shape) and eta (the characteristic life) from the times between failures.

- 5 Lifecycle stage
Early life, useful life, wear-out or end of life from beta, the trend, availability and the corrective share.

- 6 Prediction and interval
Next failure from the last one plus the median life; the preventive interval at the B-life for the criticality; a flag when the next task falls after the predicted failure.

Each step is computed from the records as they stand. The engineer confirms before a plan changes; a change to the interval of a qualified asset goes through change control.

## Calibration belongs in the same model
A calibration that finds a reading out of tolerance is a failure of the measurement, and in GMP it carries a heavier consequence than a seized bearing: everything the instrument measured since its last good calibration was measured wrong. The calibration module treats it as the regulations do. The reading is judged against the approved tolerance as it is entered, in the point's own unit or a convertible one. An as-found reading out of tolerance can be recorded as a pass after adjustment or a fail, never as a pass. The event opens the impact assessment, which the AI drafts from the readings, the requirement, the earlier calibrations and the linked records: the period at risk, the worst error, what the asset measured, the effect, the recommended decision. A person other than the one who recorded the calibration takes the decision and signs it with their password, and the investigation is linked where a result, a batch or a validated state may be affected.
Two signatures make the calibration record inspection-ready. The second-person review of each calibration checks the readings against the certificate, the standards' certificates and their expiry, and the result, and signs Reviewed, or returns it with what must be corrected. Until it is reviewed, the calibration does not count: the asset shows awaiting review, and its due date rests on the last reviewed one. QA's signature on the out-of-tolerance decision is the second. Both go to the record's signatures with the record's hash at the time, under 21 CFR Part 11, without locking a record that must stay open for the asset's life.
On a linked equipment file, each of those events counts toward the maintenance asset's reliability. A pressure transmitter that drifts out of tolerance every second calibration is telling the engineer the same thing a pump that trips every second month is: the asset is wearing, and the interval, the part or the asset should change.
What happens to 100 calibrations
- Calibrations recordedreadings judged as entered100
- Reviewed and signed by a second personthe rest returned and re-recorded96
- In tolerance as founddue date set from the interval88
- Out of tolerance as foundimpact assessment opened12
- Impact assessed, QA decision signednone left without a decision12
- Investigation openeda result or batch may be affected4
- Interval or asset changedthrough change control3

An illustrative model of a mid-sized site, not a measurement. Each stage is a count of calibrations. The losses a connected record removes are at review (nothing counts until a second person signs) and at the out-of-tolerance decision (nothing closes until QA signs with the impact assessed).

## From prediction to the maintenance plan: alerts, advisory and change control
A prediction that stays on a dashboard changes nothing. Two alert rules carry it to the people who act. When an asset's failure history predicts the next failure within the rule's lead time, the asset's owner is told, with the figures and the advice. When the asset's next preventive task is scheduled after the predicted failure, or no task is scheduled, the rule says so, with the recommended interval. Both are deterministic: they read the same model as the dashboard, every quarter hour, and raise one alert per asset until it is resolved.
The AI advisory reads the fleet's figures and the failure codes and drafts what a reliability engineer would propose: the preventive plan changes with the number of days and the task, the replacement and overhaul cases with the business case, the spares to stock and the training to give from the failure codes and repair times, and the questions the records leave open, such as the operating hours nobody recorded. It is decision support and it is labelled so. Nothing changes until the engineer takes it to the asset's plan with a reason, and to change control where the qualified state is affected.
That last step matters for the inspector. An interval shortened because the Weibull shape rose above 1.3 is a defensible decision only if the record shows the figures, the reason, who decided and when, and the requalification where it was needed. On the connected record graph it does, because the plan change is an audited edit on the asset, the change control links to it, and the next calibration and the next work order carry the new interval forward.
The predictive maintenance loop
- 1 . RecordWork orders with downtime and labour; calibrations with readings reviewed and signed; operating hours or counters.
- 2 . ComputeMTBF, MTTR, availability, trend and the Weibull fit per asset; the lifecycle stage; the predicted failure; the recommended interval.
- 3 . AlertAssets predicted to fail within the lead time, and preventive tasks that fall after the predicted failure.
- 4 . AdviseThe AI drafts plan changes, replacement cases, spares and training, and the questions the records leave open.
- 5 . DecideThe engineer changes the plan with a reason, raises the change control, budgets the replacement.
- 6 . VerifyThe next work orders and calibrations show whether the failures stopped; the model reads them on the next turn.
- The loop runs on the records the site already keeps. No sensor is required to start; sensors make the operating time exact.

Each turn of the loop reads the records the previous turn produced. The engineer decides at the plan step; the loop never changes a plan on its own.

## What changes for the site
The intervals stop being folklore. Each asset's interval has a rationale the records support, and the rationale is updated when the records change. The reactive share falls, because wear-out assets get their intervals shortened before they fail rather than after, and the useful-life assets stop being over-maintained on a schedule copied from a leaflet. The replacement cases arrive with figures, which is what finance asks for.
The calibration programme stops being a filing exercise. Due dates are kept by the system from reviewed calibrations, out-of-tolerance results cannot be closed without an assessed and signed decision, and the inspector follows the chain from the certificate to the batch the instrument measured without a spreadsheet.
The figure is an illustrative model of where the hours go before and after, not a measurement. The point is where the hours move: out of gathering and firefighting, into the work that should take the time, which is deciding what to do with the asset.
Where the maintenance engineer's month goes
- Gathering data: disconnectedthree exports, one spreadsheet24 h
- Gathering data: connectedthe dashboard, as it stands2 h
- Firefighting breakdowns: disconnectedreactive share above 60 %36 h
- Firefighting breakdowns: connectedwear-out caught early18 h
- Deciding intervals and replacements: boththe work that should take the time16 h
- Writing the change control: disconnectedassembled from emails8 h
- Writing the change control: connecteddrafted, cited, final-reviewed3 h

An illustrative model of a typical site, not a measurement. Hours per month by activity, disconnected systems first, connected record graph with the reliability model second.

## Common questions

### Does predictive maintenance need sensors or an IoT platform?

No. The model described here reads the work orders, their downtime and labour, the calibration results and the operating hours a site already records. Sensors and counters make the operating time exact and let the model work in operating hours rather than calendar days; they are an improvement, not a prerequisite.

### Can the model change a preventive interval on its own?

No. It computes the figures, predicts the next failure, recommends an interval and raises alerts. The maintenance engineer changes the plan with a reason, and a change to the interval of a qualified asset goes through change control and, where needed, requalification under EU GMP Annex 15. The AI advisory is a draft labelled as such.

### How many failures does the Weibull fit need?

Four or more on one asset, so the times between failures can be ranked. Until then the model reports the MTBF and the trend and recommends the interval as a fraction of the MTBF by criticality. With the fit, the interval targets the B-life for the criticality: B5 for Critical assets, B10 for High, B20 for Medium, B30 for Low.

### Why does a calibration need a second-person review and a QA signature?

21 CFR 211.68 and 211.160(b)(4) require records of calibration checks and remedial action when limits are not met, and FDA's data integrity guidance expects records to be reviewed. A second person checks the readings against the certificate and signs Reviewed before the calibration counts for the asset's service; when a reading is out of tolerance, QA signs the decision on the impact to what was measured. Both signatures carry the signer, the meaning and the record's hash at the time.

## Sources
- [21 CFR 211.67, Equipment cleaning and maintenance (eCFR, current)](https://www.ecfr.gov/current/title-21/section-211.67)
- [21 CFR 211.68, Automatic, mechanical, and electronic equipment (eCFR, current)](https://www.ecfr.gov/current/title-21/section-211.68)
- [21 CFR 211.160, General requirements for laboratory controls (eCFR, current)](https://www.ecfr.gov/current/title-21/section-211.160)
- [21 CFR 820.70, Production and process controls (eCFR, current)](https://www.ecfr.gov/current/title-21/section-820.70)
- [21 CFR Part 820, Quality System Regulation (eCFR, current)](https://www.ecfr.gov/current/title-21/part-820)
- [EudraLex Volume 4, EU Guidelines for Good Manufacturing Practice (Chapter 3 and Annex 15)](https://health.ec.europa.eu/medicinal-products/eudralex/eudralex-volume-4_en)
- [FDA, Data Integrity and Compliance With Drug CGMP: Questions and Answers, guidance for industry (2018)](https://www.fda.gov/media/119267/download)
- [FDA, Part 11, Electronic Records; Electronic Signatures: Scope and Application](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/part-11-electronic-records-electronic-signatures-scope-and-application)
- [FDA, Q9(R1) Quality Risk Management](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/q9r1-quality-risk-management)
- [ISPE, GAMP 5 Guide, Second Edition](https://ispe.org/publications/guidance-documents/gamp-5-guide-2nd-edition)
This article is general information for quality and regulatory professionals. It is not regulatory or legal advice. Check the source documents for their current status before acting on them.

## Topics
- CMMS
- Calibration
- Reliability
- Equipment lifecycle

## How QLogix helps
- [QMS digitization](https://qlogix.ai/services/qms-digitization)
- [Connected quality data](https://qlogix.ai/services/connected-quality-data)
- [Agentic workflow automation](https://qlogix.ai/services/agentic-workflow-automation)

[All articles](https://qlogix.ai/blog)

## More from the blog
- [AI in quality · 14 October 2026Connecting LIMS to investigations and CAPA: how AI speeds up OOS investigations and strengthens complianceMost OOS investigations lose their first days to gathering: the result, the raw data, the instrument's calibration, the standard's expiry, the analyst's history, the trend. When the LIMS, the investigation and the CAPA are one connected record graph, that evidence arrives with the result, AI reads it and drafts the hypotheses, and the effectiveness check is proven by the results that follow. The analyst, the investigator and QA still decide and sign.](https://qlogix.ai/blog/lims-investigation-capa-integration-ai)
- [AI in quality · 14 October 2026AI in deviation investigations: faster, more consistent root cause analysisAI is most useful in an investigation at the start, where time is lost: it drafts the problem statement, the containment check, the 5 Whys and a fishbone whose causes are testable hypotheses, each quoting the write-up or an attached document. The investigator then tests, decides and signs. Nothing the AI writes becomes the root cause until a qualified person says so.](https://qlogix.ai/blog/ai-deviation-investigation-root-cause-analysis)
- [AI in quality · 14 October 2026AI-assisted CAPA: from root cause to a verified effectiveness checkA CAPA fails for structural reasons: actions that restate the symptom, no owner or date, an effectiveness check invented at closure, and no view across records to see the same cause coming back. AI helps most before the plan is approved: proposing actions that address the confirmed cause, drafting the effectiveness criterion up front, and matching the CAPA against earlier records. People approve, implement, verify and close.](https://qlogix.ai/blog/ai-capa-management-effectiveness)

## Talk it through with someone who has run a plant
Bring your backlog, your systems map or one investigation that keeps coming back. We will tell you plainly what we would do and what we would leave alone.
[Request a meeting](https://qlogix.ai/contact)[Email us](mailto:contact@qlogix.ai)

In the first meeting we cover
- Your current QMS and how it connects to MES, LIMS and ERP
- Where deviations, CAPAs and change controls slow down
- Whether a readiness assessment makes sense, and what it would cover

---
Source: https://qlogix.ai/blog/predictive-maintenance-cmms-calibration-gmp
QLogix · contact@qlogix.ai · +17146639012