Where OEE Came From and Why It Travels Badly

Seiichi Nakajima introduced Total Productive Maintenance and the loss framework behind OEE at the Japan Institute of Plant Maintenance in the early 1970s. The framework organizes equipment loss into six categories: equipment failure, setup and adjustment, idling and minor stoppages, reduced speed, defects and rework, and reduced yield at start-up.1 Those six categories map onto three factors that are multiplied together. Availability captures the first two. Performance captures the next two. Quality captures the last two.

The framework earned its reputation honestly. It gave maintenance and operations a shared language, and it made the point that a machine idle for a shift and a machine running at half speed are both losing the same thing. ISO later codified the definitions. ISO 22400-2 specifies key performance indicators for manufacturing operations management, including OEE, with a defined time model that separates planned busy time from actual unit busy time and actual unit downtime, so that two sites working from the same facts should arrive at the same number.2 Academic work has since examined how consistently the standard matches the way OEE is described and used in the literature, and found meaningful variation in how practitioners define the inputs.3

The trouble is not the arithmetic. The trouble is that the arithmetic carries four assumptions that hold in an automotive stamping plant and do not hold in a pharmaceutical packaging hall.

Assumption one: a rejected unit is a failure

In discrete manufacturing, a part rejected at final inspection represents a process that produced something wrong. In pharma packaging, the line is instrumented with checkweighers, label verification cameras, seal integrity checks, torque monitors, code readers and tamper-evidence sensors, and their entire purpose is to remove units. A packaging line that rejects nothing over a long run either has no variation at all or has controls that are not challenging anything.

Assumption two: every stop is waste

Line clearance between products is not a stop that a better plant would eliminate. Under 21 CFR 211.130(e), packaging and labeling facilities must be inspected immediately before use to make sure that all drug products have been removed from previous operations and that all packaging and labeling materials not suitable for the next operation have been removed, with the inspection documented in the batch production record.4 That is a required activity with a required record. Counting the minutes as availability loss means counting compliance as waste.

Assumption three: the equipment has a known top speed

A packaging line typically runs many presentations: different carton sizes, different leaflet folds, bottle counts, blister formats, cold-form versus push-through foil, different label stocks. The line has a different achievable rate for each of them. The OEM nameplate rate was quoted for one format under favorable conditions. Performance measured against a single unvalidated maximum tells you almost nothing about whether a given run went well.

Assumption four: the three factors carry comparable information

OEE multiplies three numbers, which implicitly treats them as equally informative. In pharma the quality factor is usually close to one and barely moves, so it contributes almost no signal while still dragging the product down. A published study of a pharmaceutical company reported availability at 78.1 percent, performance at 40.0 percent and quality at 99.4 percent, producing an OEE of 31.7 percent.5 Everything worth acting on in that picture sits in performance. The composite number reports a single figure and forces the reader to go find out which factor caused it.

1971 Year Nakajima introduced TPM and the six big losses at the Japan Institute of Plant Maintenance1
34 Core KPIs defined in ISO 22400-2, the standard that specifies the OEE time model2
99.4% Quality factor in a published pharmaceutical OEE study where performance was 40.0%5

None of this makes OEE useless. It makes OEE a metric that has to be adapted before it is deployed, and adapted deliberately rather than by whatever default the historian vendor happened to configure. The rest of this article works through what the adaptation looks like, factor by factor.

The Quality Factor Is Where OEE Goes Wrong

Start here, because this is the point that matters most and the one most often left unexamined.

The OEE quality factor is good units divided by total units started. Nothing in the definition asks why a unit was rejected or who rejected it. On a pharmaceutical packaging line, that omission is fatal, because the great majority of rejected units are removed by automated in-line controls that exist specifically to remove them.

A reject is usually the control working, not the process failing

Walk a packaging line and count the reject stations. A checkweigher removes bottles outside the fill weight window. A vision system verifies the printed lot number and expiry date and removes cartons where the characters cannot be read with confidence. A code reader grades the 2D data matrix used for serialization and removes any carton whose code will not scan reliably downstream. A seal integrity sensor removes blisters with a suspect seal. A leaflet detector removes cartons where the insert did not go in. Each of these is a qualified control, and each of them produces rejects during normal operation.

When one of these controls removes a unit, three things are simultaneously true. A defective or unverifiable unit did not reach a patient, which is a good outcome. The control demonstrated that it is functioning, which is evidence you want. And something upstream produced a unit that needed removing, which is a process signal worth chasing. Only the third of those is a quality loss, and it is a loss at the level of the upstream cause, not at the level of the reject station.

The standard OEE quality factor collapses all three into one arithmetic penalty.

The perverse incentive is real and it is specific. If the quality factor is in the OEE number, and the OEE number is in a site scorecard or a bonus target, then every tightening of an inspection threshold makes the site look worse and every loosening makes it look better. Nobody has to intend anything improper for this to bend behavior. It is enough that a threshold change becomes a conversation about the metric rather than a conversation about detection capability. Inspection setpoints are configuration under change control. The reason recorded on that change is discoverable, and “to improve OEE” is not a reason anyone wants an inspector to read.

Separate the reject stream into three classes

The fix is not to abandon the quality factor. It is to stop treating the reject stream as one thing. Three classes behave differently, are owned by different people, and need different responses.

Reject classWhat it isWho owns itWhere it belongs in the metrics
True positive control reject A defective or non-conforming unit correctly removed by an in-line control The upstream process or material that created the defect Defect Generation Rate, attributed to the upstream cause. Not a fault of the reject station.
False positive control reject A conforming unit removed because the inspection system could not verify it (poor contrast, print variation, sensor drift, marginal setpoint) The inspection system and its configuration, plus the material supplier where print or substrate variation is the driver Inspection False Reject Rate, tracked as a measurement system property. Never a process quality loss.
True end-of-line quality loss Start-up and setup scrap, units lost at reconciliation, material quarantined or rejected by QA, product rejected on a deviation Production and quality jointly The OEE quality factor, if you keep one. This is the only class the classical definition describes correctly.

Once the stream is split, the picture becomes readable. A rising true positive rate says a real problem is developing upstream and the control is catching it. A rising false positive rate says the inspection system is drifting, or a supplier changed something about the substrate, or a setpoint is too aggressive for the current format. A rising end-of-line loss says setup or reconciliation is degrading. These are three different investigations. Blending them into one percentage guarantees that none of them gets started.

False rejects are a measurement system problem, and they are large

Practitioners consistently underestimate how much of a reject stream can be false. A documented pharmaceutical line implementation reported that the system being replaced was rejecting 25 percent of good packages, and that after moving to optical character recognition and verification the false reject rate fell to 0.5 percent, with a corresponding improvement in first pass yield and OEE.6 A quarter of good product being pushed into the reject bin is not an exotic scenario. It happens whenever an inspection recipe is tuned for a worst-case format and then run on everything.

The uncomfortable part is that you cannot know your false reject rate from the line data. The reject counter counts rejects. It does not know whether the unit was defective. The only way to find out is to look at rejected units and adjudicate them, on a defined sampling plan, with the results recorded. This is ordinary measurement system analysis applied under GMP documentation discipline, and very few packaging halls do it routinely.

Practical routine. Pull a defined sample from each reject station per shift or per batch, adjudicate each unit against the same acceptance criteria the control applies, and record the split between correctly rejected and incorrectly rejected. Trend the false reject rate by station, by product and by material lot. Two things fall out almost immediately: which stations are miscalibrated, and which packaging component suppliers are driving the variation. Both are actionable. Neither is visible in OEE.

What the regulation asks for

It is worth being precise about the regulatory frame, because the pressure to loosen inspection tends to be argued in operational language while the constraint is regulatory. In-process control samples and testing are required to assure batch uniformity and integrity, with control procedures established to monitor output and validate the performance of processes that may cause variability.7 Labeling issuance and reconciliation are separately required, and discrepancies between the quantity of drug product finished and the quantity of labeling issued that fall outside narrow preset limits have to be investigated.8 Production record review requires investigation of any unexplained discrepancy or failure of a batch to meet specifications, with the investigation extended to other batches that may be associated.9

Read together, these establish that reject and reconciliation data are quality records with investigation obligations attached, not simply operational counts feeding a dashboard. A metric design that treats them as operational counts is not just analytically weak. Without anyone intending it, it puts a performance incentive on top of a regulatory obligation.

What good looks like. The site publishes two quality-side numbers instead of one. Defect Generation Rate counts defective units per million produced, attributed to the upstream cause that made them. Inspection False Reject Rate counts conforming units incorrectly removed, attributed to the control system and its inputs. Both are expected to fall over time. Tightening a threshold moves the second number up temporarily and the first number down permanently, and both movements are visible and explainable. Nobody has to argue about whether a better control made the plant look worse.

Availability: Line Clearance Is a Requirement, Not a Loss

Availability is the ratio of time the line ran to the time it was scheduled to run. The classical treatment puts changeover in the numerator of loss, because in a discrete plant a setup is a period the machine could have been making parts and was not. In pharma packaging, that framing collapses the moment you look at what a changeover contains.

What a pharma packaging changeover involves

A product-to-product changeover on a packaging line is not a tool swap. It is a sequence with regulatory obligations attached at several points.

1

Removal and reconciliation of the previous product

All remaining product, cartons, leaflets, labels and printed components from the previous batch are removed and accounted for. Labeling reconciliation happens here, and discrepancies trigger investigation.

2

Cleaning to the required standard

Cleaning is performed against a validated procedure, with the level determined by product changeover risk. This is not discretionary time and cannot be compressed by working faster.

3

Documented line clearance inspection

An inspection, immediately before use, to confirm that the line and area are free of previous materials, with the result recorded in the batch production record as required by 21 CFR 211.130(e).

4

Format changeover and setup

Change parts fitted, guides set, sensors and inspection recipes selected for the new format, coder and serialization job loaded. This is the part that most resembles a discrete-manufacturing setup.

5

Setup verification and first-unit checks

Verification that the correct components are present, the correct print is applied, the inspection systems detect a known bad sample, and the first units meet requirements before the run is released.

Steps one, three and five are required activities that produce required records. Steps two and four contain real waste and real opportunity. An availability figure that lumps the whole block into downtime is reporting a compliance obligation as a loss, and it invites the wrong improvement conversation.

How to classify it

The ISO 22400 time model already gives you the mechanism: planned busy time excludes planned downtime, and OEE is calculated against planned busy time rather than calendar time.2 Scheduled line clearance and cleaning belong in planned downtime, with their own category and their own visible number. They should never be absorbed without notice into either the numerator or an unlabeled bucket.

Three rules make this work in practice.

  • Planned versus unplanned is an explicit flag, not an inference from the reason code name. A cleaning event that was scheduled and a cleaning event triggered by a spill are the same activity with completely different meaning. If the only field is a reason code called “Cleaning”, the data cannot tell them apart.
  • GMP-required versus discretionary is a second explicit flag. This is the flag that stops line clearance from being targeted as waste. It also makes it possible to report an honest schedule utilization figure that shows how much of the shift the regulation consumed.
  • Changeover duration gets its own metric with its own standard. Not a percentage inside availability. A duration, by changeover type, against a defined standard time, with the five-step breakdown above so the improvable part is separable from the required part.

What a real changeover improvement looks like

Single-minute exchange of dies work has been applied to pharmaceutical changeovers with documented results, and the technique is sound as long as it is applied to the right steps.10 The distinction is straightforward once stated.

GENUINE

Converting internal work to external

Kitting change parts and staging them at the line before the run ends. Pre-staging printed components against a verified pick list. Preparing and pre-approving the batch documentation package. Loading the next serialization job in advance. None of these remove a required step. They move preparation off the critical path.

GENUINE

Parallelizing across line modules

Clearing and re-setting the cartoner while the bottling module is still finishing, where segregation allows it and the clearance record can still be attributed correctly. This needs a documented approach so the clearance evidence stays defensible.

GENUINE

Removing the wait, not the check

A large share of changeover duration is often waiting for a second person to arrive and verify. Scheduling the verifier against the changeover plan rather than paging when the line is already stopped removes waiting time without removing the verification.

GENUINE

Engineering the format change out

Tool-free change parts, keyed guides that cannot be set wrong, recipe-driven setpoints replacing manual adjustment. This reduces both the duration and the setup verification burden, because there is less to verify when fewer settings are manual.

NOT GENUINE

Compressing or combining the clearance

Performing one clearance for two consecutive products, shortening the inspection, or narrowing what is inspected. This changes the control, not the process, and it is the kind of change an inspector will find in the record.

NOT GENUINE

Signing ahead of the check

Recording verification before it is performed, or having the operator verify their own setup. This is a data integrity failure regardless of what it does to the changeover time. Contemporaneous recording is a defining expectation of GxP data.11

A test worth applying before any changeover improvement is approved. Ask whether the change reduces the time a required activity takes, or reduces the extent of the required activity itself. The first is improvement. The second is a control change and belongs in change control with a quality assessment, not in a continuous improvement report. Sites that skip this distinction end up with a shorter changeover and a finding.

Performance and the Rated Speed Nobody Validated

Performance compares the units produced to the units that could have been produced at the ideal rate over the same running time. The entire factor rests on the denominator, and in most packaging halls the denominator is a number that came from a sales specification and was never confirmed on site.

The nameplate rate is not the ideal rate

The OEM rated speed is quoted for a defined format under favorable conditions with specified materials. It is a commercial number as much as a technical one, and it is written to be achievable in a factory acceptance test rather than sustained across a shift with real components. Setting the ideal cycle time too fast means performance is permanently below 100 percent no matter how well the line runs, which teaches everyone that the number is unachievable and can be ignored. Setting it too slow inflates performance and hides real speed loss.12 Neither error announces itself.

There is a second and larger problem. A packaging line does not have one ideal rate. It has one per product and format. The achievable rate for a 30-count bottle of a free-flowing tablet is different from the achievable rate for a 90-count bottle with a desiccant, and both are different from a blister format with a folded leaflet and a cold-form foil. Measuring all of them against a single number does two things at once: it makes every line look bad, and it hides the variation between products and shifts that is the only thing you could act on.

Basis for the ideal rateWhat it producesUse it?
OEM nameplate rate, one number for the line Performance permanently depressed, all products look equally poor, no signal about which product or shift is the problem No. Keep it as a reference for capacity planning only.
Design rate from the URS or capacity model Useful for investment decisions, disconnected from what the line does today with current components No, not as the performance denominator.
Best-ever recorded rate for the product Anchored on a single lucky run, often with an unrepresentative material lot. Unreachable and demoralizing. No. Single-point maxima are noise.
Demonstrated sustained rate, per product and format A rate the line has held for a defined continuous window under normal conditions, derived from the data and recorded with its basis Yes. Requalify when the format, components or equipment change.

Measure at the constraint, not everywhere

A packaging line is a set of connected machines with buffers between them. The line output is set by the constraint machine plus the behavior of the buffers around it. Measuring performance at a non-constraint machine tells you how often that machine was starved or blocked by something else, which is real information but is not a performance problem belonging to that machine. Reporting it as one invites operators to speed up a machine that then simply fills its downstream buffer and stops.

The workable approach is to measure rate attainment at two points: the constraint machine, and the line exit. The gap between them is the loss to starving, blocking and rejects between the two, and it is usually more interesting than either number on its own. Adding per-machine state data underneath lets you find which upstream machine is doing the starving without turning every machine into a scored entity.

Deliberate slow running is a decision, not a loss

Packaging lines are frequently run below their achievable rate on purpose. A line may be slowed to reduce carton misfeeds on a marginal component lot, to keep a fragile format intact, to hold a code reader within its reliable read window, or because a downstream manual operation cannot keep up. Every one of these is a reasonable decision, and every one of them appears in OEE as unexplained speed loss.

Make deliberate rate reduction an explicit, recorded state. Give it a reason code with an owner and an expected end condition. A line running at 80 percent of demonstrated rate because a component lot is out of specification should show as a material issue with a supplier owner, not as a performance loss with an operations owner. Without this, the single most common driver of speed loss in packaging (component variation) is invisible in the metric that is supposed to find it.

Where the Data Comes From

Any discussion of what to measure runs into the question of what can be collected in practice, and at what confidence. On a packaging line the answer splits cleanly into two categories with very different reliability.

AUTOMATIC

Machine state and timing

Running, stopped, starved, blocked, in setup. Taken from the PLC or line control system with accurate timestamps. Reliable, high resolution, and available whether or not anyone is paying attention.

AUTOMATIC

Counts

Infeed count, good count at the exit, reject count per station. Taken from encoders, photoeyes and reject confirm sensors. Accurate, though the definition of what each counter counts needs to be written down and agreed, because counters placed at different points do not reconcile by themselves.

AUTOMATIC

Fault and alarm codes

The machine’s own diagnostic code for what tripped. Useful, but it names the sensor that detected a condition, not the cause. A jam detected at the carton infeed can originate three stations upstream.

AUTOMATIC

Serialization and coding events

Code applied, code read, code graded, code commissioned or decommissioned. Rich, timestamped, and already retained for regulatory reasons, which makes it usable for line analysis at no additional collection effort.

OPERATOR-DEPENDENT

Reason for a stoppage

The system knows the line stopped and for how long. It does not know why. Someone has to say. Everything downstream of this classification depends on it being both accurate and timely.

OPERATOR-DEPENDENT

Which fault was the real one

A single event often trips several alarms in sequence. The machine records all of them. Only a person can say which one started it, and only if asked at the time rather than at the end of the shift.

OPERATOR-DEPENDENT

Planned versus unplanned

The line control system cannot tell a scheduled clearance from an unscheduled one. Unless the schedule is integrated, this is a human classification, and it determines whether the time lands in planned downtime or in availability loss.

OPERATOR-DEPENDENT

True versus false reject

Entirely invisible to automation. The reject station knows it rejected something. Only adjudication of the rejected units establishes whether it should have.

Classification latency is a data quality problem

The single most common way OEE data becomes unusable is delayed classification. A reason entered at the end of a shift for a stop that happened six hours earlier is a reconstruction, not a record. The operator remembers the big events and guesses at the small ones, which systematically biases the data toward large sporadic events and away from chronic small stops. That bias then drives the improvement program toward the wrong problem, which is discussed in more detail below.

The practical requirements are unglamorous and they matter more than the analytics layer sitting on top. Classification happens at the machine, on a device the operator can reach with gloves on. The list of choices is short enough to select in a few seconds. Unclassified stops are visible in real time to the line lead rather than discovered at month end. And the system enforces a maximum classification age, after which the stop is marked as unclassified rather than backfilled without notice.

Decide deliberately whether the system is a GMP system

This question gets deferred and should not be. An OEE system that only informs continuous improvement is not a GMP record system. The moment its output feeds an annual product review, supports a deviation decision, provides the reject counts used in reconciliation, or is cited in an investigation, it is producing regulated data and the expectations follow.

EU GMP Annex 11 sets requirements for computerized systems used in GMP activities, including validation proportionate to risk, access control, risk-based consideration of audit trails for GMP-relevant changes and deletions, and checks on the accuracy of data entered manually or transferred electronically.13 The MHRA data integrity guidance defines the record expectations in more detail, including what counts as raw data and the expectation that records are created at the time of the activity.11 Neither of these is onerous for a line performance system designed with them in mind. Both are painful to retrofit onto a system that was deployed as an operations dashboard and then, without a formal decision, started feeding quality documents.

Write the boundary down. A single page stating which outputs of the line performance system are used for GMP decisions, which are used for improvement only, and what controls apply to each. Then hold the line. The failure mode is not a deliberate decision to use unvalidated data in a regulated context. It is a quality colleague finding a convenient report and pasting a number from it into an investigation, eighteen months after anyone remembered the system was out of scope.

The Reason Code Taxonomy Decides Whether Any of This Is Useful

Of everything discussed here, the reason code taxonomy has the largest effect on whether line performance data leads to any action, and it receives the least design attention. It is usually configured during system commissioning by whoever was available, copied from a vendor template, and never revisited.

How to recognize a taxonomy that has failed

  • The “Other” bucket is large. If more than a small share of downtime minutes is recorded as Other, Miscellaneous or Unclassified, the taxonomy does not describe what happens on the line, and no amount of analysis will recover the information.
  • Codes name a machine instead of a failure. “Cartoner stop” tells you where the line stopped. It does not tell you what to fix. Location and cause are two different fields.
  • Codes describe what the operator did. “Cleared jam” and “Reset machine” record the response, not the condition. Every fault ends with someone clearing or resetting something.
  • There are hundreds of codes. Beyond roughly forty choices, operators stop searching and select from the top of the list or the most recently used. The data then reflects list position rather than reality.
  • Cause and consequence are mixed. “Waiting for materials” and “Label web break” belong to different owners and different countermeasures, and a flat list gives no way to say so.
  • Nobody has ever acted on a specific code. If a code has never appeared in an improvement action or a maintenance change, it is collecting keystrokes and returning nothing.

Design rules that hold up

1

One code, one countermeasure owner

Every code should map to a single function that would own the fix: maintenance, engineering, production, quality, planning, or supplier quality. If two owners could plausibly claim a code, split it. If none can, delete it.

2

Shallow at entry, deep behind it

The operator picks from eight to twelve top-level categories, then from a short second level within the chosen category. Two quick selections beat one long scroll, and the resulting data has a usable hierarchy.

3

Three fields, not one string

Location (which module), cause (what condition), and consequence (what the line did) are separate. This is what makes it possible to ask whether carton misfeeds are concentrated on one format, or whether one module produces a disproportionate share of a common cause.

4

Planned and GMP-required are flags, not code names

Any activity can be planned or unplanned. Any stop can be required or not. Encoding this in the code name creates duplicate codes and inconsistent selection. Two boolean fields solve it cleanly and make honest reporting straightforward.

5

Material and component causes are first-class

Component variation is one of the largest real drivers of stops on packaging lines and is almost always buried under machine codes. Give it explicit codes with a supplier field so the data can reach supplier quality instead of dying in a maintenance report.

6

Review the taxonomy on a schedule and retire dead codes

Quarterly, list every code with its share of downtime minutes and every improvement action taken against it. Codes with volume and no action indicate a stalled problem. Codes with no volume indicate a code nobody uses. Both need a decision.

Treat the taxonomy as configuration, and version it. Changing reason codes changes what the historical data means. A category renamed or merged breaks year-over-year comparison without any warning, and analysts will spend weeks explaining a trend that is an artifact of a configuration change. Version the taxonomy, date the changes, and keep a mapping from old codes to new ones. If the system is in GMP scope, this is a change control obligation anyway.

Coverage is itself a metric

The most useful early indicator of whether a line performance program is working is not availability or OEE. It is coverage: the percentage of downtime minutes that carry an actionable, non-Other reason code, classified within the shift in which they occurred. A site with 60 percent coverage does not have a performance problem it can analyze. It has a data collection problem it has to fix first, and reporting OEE from that data will produce confident conclusions about the 60 percent that happened to be recorded.

Chronic Small Stops Versus a Few Large Events

Two packaging lines can report identical availability and have nothing in common. One lost four hours to a single servo failure. The other lost four hours to roughly six hundred stops of forty seconds each. The aggregate figure is the same and the correct response is completely different. This is the most consequential analysis question in line performance, and it is invisible in the headline number.

Why small stops hide

Short stoppages disappear for two structural reasons. First, most systems apply a duration threshold below which a stop is not treated as downtime at all, so brief stops fall into speed loss inside the performance factor rather than into availability. Second, brief stops are rarely classified, because clearing a jam and getting the line moving takes less time than logging what happened. The result is that the loss category most likely to dominate real packaging line performance is the one least likely to appear in the data with a reason attached.

Nakajima’s original framework already separated idling and minor stoppages from equipment failure precisely because they behave differently and demand different countermeasures.1 Practice tends to lose that distinction as soon as the numbers are rolled up.

Look at the distribution, not the total

The diagnostic is simple and rarely done. For a given line and period, plot two things: the count of stops by duration band, and the total minutes lost by duration band. The shape tells you immediately which problem you have.

Chronic small stopsSporadic large events
Signature in the data High count, short duration, minutes concentrated in the under-five-minute bands, often clustered at one or two modules Low count, long duration, a handful of events carrying most of the lost minutes
Typical causes in packaging Carton misfeeds, label web breaks and splices, leaflet folding faults, code reader no-reads, reject gate clearing, bottle infeed jams, component dimensional variation Servo or drive failure, sealing station failure, utilities interruption, serialization system outage, a deviation that stops the line, a shortage of a printed component
Right response Engineering the condition out: change part tolerances, guide and transfer geometry, material specification tightened with the supplier, machine accessibility so clearing is quick, condition monitoring of format parts Maintenance strategy and reliability work: failure investigation, spares availability, preventive and predictive maintenance intervals, redundancy or bypass design, supply planning
Wrong response and what happens Running a formal root cause investigation on each jam. You get a stack of reports and the jams continue, because the cause is a tolerance stack, not an event. Running an improvement workshop on a failure that occurs twice a year. There is nothing to observe and no data to work from, and the team concludes the method does not work.
Who usually owns it Engineering and supplier quality, with production input Maintenance and reliability, with planning input

The material connection nobody looks at

A large share of chronic small stops on packaging lines traces back to packaging component variation rather than to the equipment. Carton caliper and cut quality, glue flap dimensions, leaflet fold accuracy and paper stiffness, label liner and adhesive behavior, foil tension and gauge, bottle neck finish consistency. Each of these can hold a machine within tolerance on one lot and drive it out on the next, with no change to the equipment at all.

This is why the component and supplier fields in the reason code taxonomy matter so much. Without them, every component-driven jam is recorded as a machine fault, the analysis points at the machine, engineering adjusts the machine to accommodate a bad lot, and the next good lot then runs worse. Sites that add a supplier and lot field to their stop records frequently find that a small number of component lots account for a disproportionate share of their small-stop minutes, and that finding leads somewhere useful.

A useful monthly view. Total lost minutes split three ways: chronic (stops under five minutes), sporadic (stops over thirty minutes), and the middle band. Each split further by whether the reason code points at equipment, material, method or people. Four numbers and a stacked bar, and it will tell a site leadership team more about where to put engineering effort than a year of OEE trend charts.

A Measurement Set That Works in a Packaging Hall

The practical recommendation is not to abolish OEE. It has a real use in capacity planning and in comparing a line against itself over time, and it is understood across the industry, which has value. The recommendation is to stop publishing the composite number on its own, to fix each factor’s definition, and to add the small number of measures that carry the information OEE loses.

MeasureDefinitionSourceOwner
Schedule utilization Running time as a share of scheduled time, with GMP-required planned downtime reported separately and visibly rather than netted out Line control system plus schedule integration Production
Unplanned downtime minutes Split into chronic (under five minutes) and sporadic (over thirty minutes) bands, by reason category Line control system plus operator classification Production and maintenance
Changeover duration By changeover type, against a defined standard, decomposed into required steps and improvable steps Line control system plus schedule and batch record Production and planning
Rate attainment Actual rate against the demonstrated sustained rate for that product and format, measured at the constraint and at the line exit Counts and machine state Engineering
Defect generation rate Non-conforming units per million produced, attributed to the upstream cause Reject counts plus reject adjudication sampling Production and quality
Inspection false reject rate Conforming units incorrectly removed, by station, product and material lot Reject adjudication sampling Engineering and supplier quality
Reconciliation variance Difference between components issued, used, rejected and returned, per the labeling accountability requirement Batch record and counts Quality
Reason code coverage Share of downtime minutes carrying an actionable code classified within the shift, plus the Other share Classification system Production leadership

Connect it to the quality system rather than running it beside one

ICH Q10 sets the expectation that a pharmaceutical quality system includes a process performance and product quality monitoring system, applies knowledge from it, and uses management review of process performance to drive continual improvement.14 Line performance data is exactly the kind of input that system is meant to consume, and the FDA process validation framework treats ongoing process verification as a continuing obligation rather than a one-time event.15

There is also a reporting connection worth understanding. FDA’s quality metrics work has centered on a small set of measures including lot acceptance rate, product quality complaint rate and invalidated out-of-specification rate, described in draft guidance and developed with industry input.1617 A site whose reject stream is correctly classified can explain its lot acceptance rate. A site whose reject stream is a single undifferentiated number cannot, and it will spend the conversation reconstructing from batch records.

Be careful what you set a target on

The last point is a governance point and it deserves to be stated plainly. Measurement changes behavior, and a target changes it faster. Putting an OEE target on a packaging line puts a target on the quality factor, and the quality factor rises when the line rejects fewer units. There is a legitimate way to reject fewer units, which is to produce fewer defective ones, and it is slow. There is an illegitimate way, which is to detect fewer of them, and it is fast.

If a single composite number has to exist for reporting purposes, remove the quality factor from it. Publish an availability and performance composite for operational performance, and publish defect generation and false reject rates separately as quality measures with their own review path. Nothing is lost analytically, and the incentive to weaken detection goes away entirely. Packaging line performance and packaging line quality are different questions with different owners, and multiplying them together was a convention from an industry where the answer to both was the same.

The broader point applies past packaging. A metric imported from another industry arrives with that industry’s assumptions embedded in its arithmetic, and those assumptions are usually invisible because they were never controversial where the metric came from. In a GMP environment, the assumption that a rejected unit represents a failure and that a stopped line represents waste are both wrong often enough to matter. Adapting the metric is not a compromise of rigor. Leaving it unadapted is.

Conclusion

OEE is worth keeping in a pharma packaging hall, but only after the definitions have been rebuilt around what happens there. The quality factor has to distinguish a control functioning correctly from a process producing defects, because a plant that tightens inspection should not see its performance number fall. Availability has to treat line clearance and documented setup verification as required planned activity with their own visible measure, not as loss. Performance has to be measured against a demonstrated, product-specific rate that the line has held, not against a vendor figure that was never confirmed on site. And all of it depends on a reason code taxonomy designed so that a person standing at a machine with thirty seconds to spare can record something an engineer can act on three weeks later.

The organizations that get value from line analytics are not the ones with the most sophisticated platform. They are the ones that decided what each number means, wrote down which decisions it is allowed to support, resisted the pull toward a single headline figure, and then looked at the distribution of stops rather than the average. That is unglamorous work and it usually takes a few months. It also tends to be the difference between a dashboard nobody trusts and a set of numbers that direct engineering effort at the right problem.

Sakara Digital works with pharma and biotech organizations building manufacturing performance measurement that holds up in a regulated environment. If you are standing up line analytics, rebuilding a reason code taxonomy, or trying to work out why your OEE number and your operators disagree about how the line is running, we are happy to have that conversation.

For Further Reading