Key takeaways
See the OEE calculation this batch variant modifies.
The classic OEE definition, availability × performance × quality, was built around a discrete production model. A bottle goes in, a finished good comes out, every X seconds. The performance number is "actual cycle time vs ideal cycle time." Slow the line by 10% and the number drops by 10%. Clean and simple.
Batch processes do not work like that. A bioreactor charges for two hours, runs for forty-eight, cleans for six, holds for two, transfers for one. There is no single "cycle time." A pasteuriser runs at a continuous speed but the input is a batch, with a defined start and end.
A coating line on a tablet press has speed variations within a batch that are by design, not loss. Force the discrete OEE formula onto these and the numbers are not just imprecise, they are misleading. Plants that benchmark against them tune for the wrong thing.
The problem is not OEE as a concept; OEE for batch is genuinely useful. The problem is the unmodified formula. A working batch OEE keeps the three components but redefines each so the math matches the process.
Discrete availability is "run time / planned production time." Both halves of that fraction are slippery in batch.
In a batch process, the same asset moves through states that are all "planned" but only some are "productive." A bioreactor's CIP cycle is planned; it is not producing. A reactor's hold time is planned; it is producing in a chemical sense but not adding throughput.
The denominator should count only states the plant considers productive, usually charge, react/run, discharge, and exclude CIP, idle, and changeover from the OEE clock the way you would exclude a planned shutdown in discrete.
Run time should be measured in productive states, not "asset on." A reactor that is running but in a hold state for hours because the downstream tank is full is not contributing throughput. That hold time should subtract from run time the same way a minor stop would in discrete. The article on production loss analysis covers the loss-class framework these states map to.
Discrete performance is straightforward: "ideal cycle time / actual cycle time." Batch processes do not have a single cycle time. They have a batch duration, a batch yield, and a batch consistency.
The adapted metric is yield-based: "actual batch yield / design batch yield." A reactor with a design yield of 950 kg producing 880 kg has a performance of 92.6%. The number is stable, repeatable, and moves with real causes, temperature excursions, reactant impurity, mechanical issues mid-batch.
For continuous-with-batch-feed processes (pasteurisers, coaters), the right performance metric is actual processing rate during productive state vs design rate. Subtle, but the difference between this and the discrete formula is what makes the number trustworthy.
Discrete quality is "good units / total units" measured in-line. Batch quality is measured at release, sometimes hours, sometimes days after the batch is produced. The OEE quality number cannot wait for release without becoming useless for daily decisions.
The working approach is two-tier. The daily OEE uses an in-process quality proxy: deviations, alarms, out-of-spec readings during the batch. The monthly OEE roll-up uses actual release data and reconciles against the proxy. Over time the proxy becomes well-calibrated against release outcomes, and the daily number stays meaningful.
The piece on manufacturing KPIs covers the broader KPI families this connects to. For pharma and food specifically, the release process is also a regulatory artefact, keep the OEE quality number separate from the audit-trail quality record to avoid creating a parallel system that drifts out of agreement.
On an adapted batch OEE formula, a well-run process typically produces:
A plant that reports composite OEE above 90% on batch is almost certainly mis-applying the formula. A plant that reports below 60% has either a real problem or a denominator that includes unproductive states it should exclude.
The diagnostic question, "which component is dragging the number?", only works after the components are correctly defined per the above. The article on root cause analysis covers how to investigate each component once the metric is right.
The other place batch processes break discrete OEE thinking is in maintenance. A reactor failure that delays a charge by four hours is not "minor mechanical", it is a batch-skip event that may cost a full day of downstream production. The loss taxonomy needs a "batch-impact" classification on top of the discrete one.
The corresponding work order management system needs to know that this asset's failures cost more in production minutes than the asset itself would suggest.
Without that classification, the maintenance prioritisation under-weights batch-process assets. The PM frequency on the reactor stays the same as it would be on an auxiliary fan. The team is surprised when the same asset keeps showing up as a quarterly loss. The connection upstream into the preventive maintenance schedule is what makes the priorities match the actual cost of failure.
The adapted batch OEE formula can be implemented in any OEE system that allows custom availability denominators and a yield-based performance calculation. The hard part is keeping it consistent across multiple lines and multiple sites, and tying the loss classes back into the work-order priorities.
Fabrico is built so the same loss taxonomy, the same asset hierarchy and the same work-order rules apply to batch and discrete lines under one platform, the formula adapts per line, the math stays comparable. To see how batch OEE looks against your live process data, book a demo .
You can, but the number will not move with real causes. Plants that do this usually end up ignoring OEE on their batch lines within a year because the number does not respond to anything they try. The cost of adapting the formula once is much lower than the cost of an unreliable metric being on the wall for two years.
Run two OEE numbers: a discrete OEE for the fill line and a batch OEE for the upstream. Reporting them together produces a misleading aggregate. Most plants find that the bottleneck is in one of the two, not both, so the two-number view is more diagnostic than a single composite.
Carefully. Most industry benchmarks for "OEE in pharma/food" are computed with inconsistent denominators, so the published numbers vary by a wide margin for the same actual performance. Benchmark against your own historical baseline and against your design yield, not against external reports.
Treating the adapted formula as a quick fix and not updating the loss taxonomy to match. The whole point of the rework is that the components move with batch-specific causes; that only works if the loss classes are batch-specific too. Half-adapted formulas are worse than the unadapted version because they signal "we have batch OEE" without delivering it.
For a single process line, 3-4 weeks to define the components and run side-by-side against the old metric. For a multi-line plant, plan on 8-12 weeks to do all the lines without losing the team. The longest pole is usually the quality-component definition, because it touches the release process.
Login
Blog
