OEE for Process Operations: What Actually Limits Throughput

Overall Equipment Effectiveness is the most quoted number on a plant report and one of the most misread. It promises a single percentage that tells you how well an asset turns time into good product. On a discrete assembly line that promise mostly holds. On a process line that dries, presses, extrudes, smelts, or ferments, the same arithmetic quietly measures the wrong things, and the number on the dashboard stops meaning what the people reading it think it means.

OEE was formalized as a manufacturing KPI in ISO 22400-2:2014, the standards-body definition of key performance indicators for manufacturing operations management, and it predates that standard by decades. A 2016 NIST study describes it plainly as a quantitative measure of the productivity of production equipment, built from availability, performance, and quality. The math is the same everywhere. The interpretation is not. What follows are the claims about OEE that cause the most damage in process operations, and why each one falls apart on contact with a real line.

"Get to 85% and you're world-class."

The eighty-five percent target is the single most repeated fact about OEE, and almost nobody who repeats it knows where it came from. It traces to Seiichi Nakajima's 1988 book that introduced Total Productive Maintenance to English-language industry, the same work NIST credits with originating the metric. According to the literature review by Muchiri and Pintelon (2008), Nakajima's world-class benchmark is the product of three component targets: availability near ninety percent, performance near ninety-five percent, and quality near ninety-nine percent. Multiply those and you land at roughly eighty-five.

Notice what that benchmark assumes. It was drawn from observation of equipment-paced discrete lines in high-performing Japanese plants, where each station has a clean ideal cycle time and a countable unit of output. A continuous reactor or a rotary dryer has neither. Its "cycle" is a flow rate in tonnes or litres per hour, its downtime includes mandatory cleaning and grade changes, and its quality is a property of a stream rather than a tally of conforming parts. Port a target built on those discrete assumptions into that world and it tells you nothing about whether the plant is healthy.

So is the benchmark wrong? Not exactly. The components are sound; the headline is the problem. The same review by Muchiri and Pintelon notes the gap between OEE in theory and OEE in practice is wide, and that real plants commonly sit well below the world-class figure. Chasing a borrowed target encourages the worst behaviour in measurement: redefining the inputs until the output looks right. Widen the definition of planned downtime, pick a gentle ideal rate, and you can manufacture a world-class score without touching the process. The number goes up; the truck leaves no fuller.

There's a second cost to the borrowed target, and it shows up in management reviews. A plant that holds itself to someone else's eighty-five learns to argue about the number instead of the loss behind it. The whole point of writing OEE into a standard like ISO 22400-2 was comparability: a line in one factory and a line in another are measured identically only if the definitions match. Two plants chasing the same headline with different accounting for cleaning, warmup, and grade changes aren't comparable at all, and the comparison does real harm when capital follows it. So drop the target worship. The useful question is never whether you've reached a borrowed percentage. It's which of the three factors is bleeding, whether that loss is trending up or down against your own history, and what the loss is actually costing in material and time. Treat eighty-five as a historical footnote, and your own last quarter as the benchmark that matters.

"OEE is one number for the whole plant."

This is the rollup fallacy, and it's seductive because executives want one figure they can carry into a board meeting. But OEE is defined at the level of a work unit or piece of equipment, not a site. ISO 22400-2 frames its effectiveness indicators around a work unit; the semiconductor industry's SEMI E79 specification for equipment productivity does the same at the level of a single tool. Average those asset-level numbers across a plant and you get a figure that describes nothing physical, because no single machine ever ran at the average.

Consider why averaging hides the truth. A process plant is a chain of stages, and throughput is governed by the slowest one. Suppose your bottleneck furnace runs at seventy percent effectiveness while four downstream units idle at high effectiveness because they're starved of feed. The plant average looks respectable, and the plant is losing money at the furnace every hour it runs. Push that average up by improving a non-constraint and you've spent maintenance hours and capital buying nothing the plant can sell. The only OEE that limits what the plant can ship is the constraint's OEE, and the constraint is usually one named unit: a dryer, a press, an extruder, a kiln, a centrifuge.

The averaging problem gets worse when the rolled-up figure becomes a target. People manage what's measured. Tie a bonus to a site-wide average and operators will, rationally, nudge the easy assets and leave the hard constraint alone, because moving the constraint is expensive and slow while polishing a non-constraint is cheap and fast. The average climbs and the bottleneck doesn't move. And the failure is invisible in exactly the report meant to catch it.

If you do want targets, set them per asset class against that asset's own demonstrated best, not against a number borrowed from another industry. A well-run rotary kiln, a packaging line, and a fermenter have nothing in common in their achievable effectiveness, and holding all three to one figure either lets the easy one coast or brands the hard one a failure for missing a goal it was never built to reach. Read each asset against its own ceiling and its own trend, flag the constraint, and let the non-constraints sit at whatever effectiveness keeps the constraint fed and the buffers healthy. A non-bottleneck idling some of the time isn't a problem to solve; it's slack the chain needs.

The honest move is to measure each significant asset on its own and read them as a chain, with the constraint flagged and watched. That's a data-collection problem before it's an analysis problem. You need per-asset state, rate, and quality at the resolution the process actually changes, not a shift-end summary typed into a spreadsheet from memory. A unit that stalls for forty seconds every few minutes never appears in a manual log, but it can halve real output. This is the work an edge telemetry and analytics platform exists to do: capture each unit's running state and instantaneous rate so the constraint reveals itself instead of being averaged away. The literature anticipated the limit. Muchiri and Pintelon trace how plant-level needs pushed OEE to evolve into broader measures such as overall factory effectiveness, precisely because one rolled-up number couldn't carry the load of a multi-stage operation.

"A high OEE means you're making more product."

OEE is a ratio, and ratios are silent about volume. Availability, in the ISO 22400-2 sense, is operating time over planned production time, and planned production time deliberately excludes scheduled downtime. So OEE says nothing about how much of the calendar you chose to run. A line scheduled for one shift a day can post a beautiful OEE while a competitor running around the clock at a lower OEE ships far more tonnage. If the question you care about is how hard the asset works across all the time you own it, OEE is the wrong instrument. The loading-inclusive measure, total effective equipment performance, divides by all calendar time and answers that question instead. They're different tools for different decisions, and confusing them leads straight to overinvesting in a line that's simply switched off half the week.

There's a subtler trap inside the performance factor. Performance compares actual rate to an ideal cycle time, and the ideal rate is something a human picks. Set it conservatively and performance looks excellent at a genuinely slow throughput. Tighten it honestly and the same line suddenly shows a performance loss it had all along. The number moved; the line didn't. Any OEE comparison across lines or across years is meaningless unless the ideal rate behind it is fixed and documented, which is one more reason the headline percentage travels so badly between sites.

And here is the failure that bites hardest on a process line: when a unit sits idle because upstream gave it nothing, naive OEE charges that idle time against the equipment as if it had broken. The paper by de Ron and Rooda (2005), whose status definitions SEMI adopted for semiconductor processes, argues the time base is the heart of the problem. Their revised measure separates equipment-dependent states from equipment-independent ones such as a lack of input. A dryer starved by a slow upstream mill isn't a dryer with an availability problem, and an OEE that blames it for the mill's shortfall will send a maintenance crew to fix a machine that was never broken.

So what does a high OEE on a non-bottleneck actually buy you? Nothing the plant can ship. Worse, it can mask the starvation that's limiting output, because a downstream asset looks busy whenever it happens to have feed, and its effectiveness is computed only over the time it was fed. Read OEE as a diagnostic of loss against intent, never as a proxy for production volume. If you want to know whether you made more product, count the product, and reconcile it against the mass balance. The dashboard percentage and the weighbridge should tell the same story; when they don't, trust the weighbridge and go find the definition that's lying. The gap usually lives in how reworked or downgraded material was booked, or in an ideal rate set years ago and never revisited as the line was debottlenecked.

"The three factors are interchangeable — just push the number up."

OEE collapses three different physical phenomena into one product, and the collapse tempts people to treat a lost point of quality as equivalent to a lost point of availability. They are not equivalent, and on a process line the difference is money. A point recovered in availability is time you didn't have before. A point recovered in quality is material, energy, and labour you already spent and then threw away. On a continuous line those sunk costs are heavy: the off-spec tonne carries its raw material, the heat or electricity already driven into it, and often the further cost of reworking or disposing of it. An availability loss costs you the margin on product you never made. A quality loss costs you that margin plus everything you poured into the scrap. Treating them as one fungible percentage erases that distinction at exactly the moment you need it to set priorities.

The decomposition matters because each factor has its own failure modes and its own fixes. Nakajima's framework groups production losses into six categories that map cleanly onto the three OEE factors, and that mapping, not the blended score, is the actual work product.

Look hard at the performance rows. Idling, minor stops, and slow running are the losses operators rarely log, because each event is small and self-clearing: a chute bridges and clears, a sensor faults and resets, a pump cavitates for a minute and recovers. Nobody writes those down. Yet on a high-rate continuous line they're often the single largest drain on output, and they're invisible without telemetry sampled faster than a person can react. A line losing a few seconds many times an hour can be down a fifth of its capacity while every shift log reads "running normally." ISO 22400-2 defines the performance ratio and the quality ratio separately for precisely this reason: so you act on the specific loss instead of the comfortable average.

Detecting minor stops is its own measurement discipline. A stop lasting a few seconds won't trip a downtime threshold built to catch breakdowns, so it hides inside the running state unless you watch the actual rate signal and treat any dip below a moving floor as a micro-stop. That means sampling motor current, line speed, or flow at a cadence of seconds rather than minutes, and tagging the cause where the instrumentation lets you (a level switch, a jam detector, a temperature excursion). Once those events are counted and ranked by cumulative lost time, the Pareto almost always surprises the people who run the line. The biggest loss is rarely the dramatic failure everyone remembers; it's usually the small recurring one nobody ever logged.

Quality deserves the same separate treatment, and process lines complicate it. First-pass quality, the fraction that meets specification without rework, is the figure that ties to cost, because reworked material consumes the line twice. A blended OEE that quietly nets reworked tonnage back into "good" output flatters quality and hides the double consumption. So resist the urge to push "the number." Push the number and you'll chase whatever factor is easiest to move, which is almost never the one costing the most. Decompose the loss, attach a cost per point to each factor using your own material and energy prices, and then act on the factor with the largest cost, not the largest gap.

"OEE is a discrete-parts metric; it doesn't fit a continuous process."

This is the comfortable myth, the one that lets a process plant skip the metric entirely and keep flying on feel. It's also wrong, but only just. OEE travels to continuous operations cleanly when you redefine its terms for flow instead of dropping the discrete template in unchanged. The standards bodies already prove the skeleton crosses domains: the effectiveness logic in ISO 22400-2 reappears in semiconductor fabs through SEMI E79, which sits beside SEMI E10 for reliability, availability, and maintainability. Different industry, same bones. What changes is the flesh on them.

Performance stops being parts per minute against an ideal cycle and becomes actual mass or volume rate against the design or nameplate rate of the unit. Quality stops being a count of good pieces and becomes the fraction of output meeting specification, measured against conforming tonnage or on-spec stream time. Availability has to account honestly for the process realities a discrete line never faces. Clean-in-place cycles, warmup and stabilization, grade transitions where the unit runs hard but makes nothing saleable, the slow ramp after a cold start where yield climbs toward spec. Each of those forces a decision: is this planned downtime, or is it a performance or quality loss while the equipment is technically running? That single accounting choice can swing the headline by ten points or more, and an undocumented choice is how two lines in one company end up uncomparable while both look fine on paper.

The hardest part is the time base, the same issue de Ron and Rooda put at the centre of their analysis. Get the state model right (productive, scheduled down, unscheduled down, starved, blocked, standby) and the metric reports the truth about a continuous unit, including the difference between a line that's broken and a line that's merely starved. Get the states wrong and OEE reports fiction with a confident decimal point, which is worse than no number, because a confident wrong number gets acted on. Building that state model is most of the engineering. It demands instrumentation on feed, rate, and quality, and a clear rule for every state the unit can occupy, decided once and applied the same way every shift.

Take a spray dryer as a worked example. Its design rate is an evaporation capacity, a mass of water removed per hour, so performance is the actual evaporation rate against that design figure and not a part count. Availability has to net out the clean-in-place cycle and the warmup needed to reach a stable outlet temperature, and you have to decide in advance whether the cold-start ramp counts as downtime or as a performance loss while the unit runs below rate. Quality becomes the fraction of powder inside the moisture and particle-size window on the first pass, with anything reprocessed charged against quality rather than quietly folded back into good output. Define those three the same way every shift and the dryer's OEE turns comparable week to week and across crews. Improvise them and it becomes a number that drifts with whoever filled in the log. That discipline, fixed definitions tied to the physics of the unit and applied without exception, is the whole game on a continuous line, and it's worth more than any benchmark you could borrow.

Not every asset needs OEE, and the metric won't fix a process that nobody acts on; a perfect number on a wall changes nothing by itself. Where it earns its keep is as a shared, decomposed language for loss, anchored to your own definitions and your own costs. But the claim that effectiveness can't be measured on a continuous line is an excuse dressed as an engineering limitation. The metric was never the problem. The discrete assumptions smuggled in alongside it were, and those you can leave at the door.