Availability vs reliability comes down to 2 different questions about the same asset. Reliability is the ability to perform as required, without failure, for a given time interval. Availability is the ability to be in a state to perform as required.
Reliability is a probability over an interval you have to state. Availability is a fraction of a clock you have to state.
Most of what you'll find online about 99.9% uptime, error budgets and 5 nines is written for IT servers. They won't help you with a compressor train.
Below are both definitions as the standards write them, what the famous availability formula measures, why your CMMS probably feeds the wrong number into it, and why no credible body publishes an availability benchmark.
Why do plants mix up availability and reliability?
Because both are maintenance objectives, and they sit right next to each other.
I summarise the objective of maintenance like this. Your equipment is available when you need it. Your equipment does what it's supposed to do. And you achieve both safely and at optimum cost.
Available when you need it is availability. Doing what it's supposed to do, without failing, is reliability.
Both numbers come out of your CMMS, off the same work orders, in the same monthly pack. So they get treated as 2 ways of saying the plant is running well.
Which one is your pack actually reporting?
What is reliability?
The international vocabulary for dependability, IEC 60050-192, defines it like this:
"ability to perform as required, without failure, for a given time interval, under given conditions"
A reliability figure needs a stated interval and stated conditions. The interval can be calendar time, operating cycles or distance, as long as the units are stated.
The vocabulary also points out that "reliability" means 2 things: the ability, and the measure of it. So when somebody tells you reliability improved, ask which one they mean.
Most maintenance dashboards report MTBF as if it were reliability. MTBF is a single average with real limitations, and our MTBF article covers when to use it and when to ditch it.
What is availability?
The same vocabulary defines availability as:
"ability to be in a state to perform as required under given conditions"
The useful material is in the notes underneath.
The first note says an item is either up or down, and that every time interval, administrative delay included, either belongs to one of those states or is excluded.
Somebody at your site decided which, and that decision sets your number.
The second note says availability depends on the combined effect of reliability, maintainability, supportability, and the maintenance and support provided.
3 of those 4 things are yours. Maintainability is design and access. Supportability is your spares, tools and documentation. "The maintenance and support provided" is your organisation on the day.
The third note goes further. Availability can also be affected by delays before anyone recognises that maintenance is needed, such as fault detection time and administrative delay.
When does your clock start: at the failure, or when somebody noticed it?
What does A = MTBF/(MTBF+MTTR) measure?
That formula usually gets printed as plain availability. Its proper name is inherent availability: the availability the design provides "under ideal conditions of operation and maintenance", with logistic and administrative delays excluded.
It's a design property. It's the availability your equipment would deliver if the spare were always on the rack, the permit were instant, and your crew stood there with the right tools the moment it broke.
There are 3 classifications:
| Classification | What downtime it counts | What it tells you |
|---|---|---|
| Inherent (Ai) | Active corrective maintenance only | What the design is capable of |
| Achieved (Aa) | Corrective plus preventive maintenance | The design plus your PM load |
| Operational (Ao) | All of the above, plus logistic and administrative delay | What you actually got |
IEC defines inherent and operational availability. Achieved availability comes from the US defence and textbook tradition, and sits between them.
Achieved availability adds preventive downtime only. Logistic and administrative delay belong in operational availability, and putting them into achieved is a common mistake.
There's a fourth clock, and it's probably the one your dashboard runs on. SMRP's Metric 2.2 defines availability as "the percentage of time that the asset is operating compared to when it is scheduled to operate". Its clock is the time the asset was scheduled to run.
SMRP calls its metric operational availability, which collides with the IEC meaning of the same 2 words.
When 2 sites compare availability, at least 4 meanings are on the table before anybody opens a work order.
Name the clock: MTTR means mean time to restoration
You can check what MTTR stands for in 30 seconds on the IEC's free online vocabulary, Electropedia. Entry 192-07-23 reads:
mean time to restoration
MTTR
DEPRECATED: mean time to repair
DEPRECATED: mean time to recovery
MTTR means mean time to restoration. "Mean time to repair" has been marked deprecated in the international vocabulary since its 2015 edition.
The measure that does mean repair is a separate entry: MRT, mean repair time. And the vocabulary is precise about what repair time contains (entry 192-07-19):
"Repair time comprises fault localization time, fault diagnosis time if not already included in fault localization time, fault correction time and function checkout time, but excludes technical, administrative and logistic delays."
A technical delay, in the vocabulary's own example, includes making the equipment safe: isolation, cooling down, grounding. Your isolation and permit to work are technical delays. They sit in the restoration number and outside the repair number.
Your CMMS almost certainly gives you work order open to work order close. That's a restoration number, and a rough one: the work order often gets raised well after the failure and closed well after the job.
Feed that into A = MTBF/(MTBF+MTTR) and you've got something closer to operational availability. Call it inherent and you've labelled a delivery number as a design number.
To make it more confusing, ISO 14224, widely used in oil and gas, uses MTTR for mean time to repair, the expansion IEC deprecated. The 4 letters stand for 2 different clocks, so say which one you mean.
2 pumps: high availability, poor reliability
The numbers below are illustrative, with a constant failure rate assumed.
2 process pumps, both required for 8,000 hours a year.
Pump A fails twice a year, and each failure takes 40 hours to restore: diagnosis, a wait for parts, a permit and the repair. MTBF is about 3,960 hours and MTTR is 40 hours, so Ai = 3,960 / 4,000 = 99.0%.
Pump B fails 20 times a year, and each failure takes 2 hours, because the spare sits on the rack and no permit is needed. MTBF is about 398 hours and MTTR is 2 hours, so Ai = 398 / 400 = 99.5%.
Pump B wins on availability.
For reliability, ask a different question. What's the probability that each pump runs a full month, 720 hours, without failing? At a constant failure rate, R(t) = e^(−t/MTBF):
| Pump A | Pump B | |
|---|---|---|
| Failures per year | 2 | 20 |
| Time to restore | 40 h | 2 h |
| Inherent availability | 99.0% | 99.5% |
| Probability of a clean month | 83% | 16% |
Pump B is half a point better on availability and about 5 times worse at getting through a month.
Break the spares assumption. Say one of Pump B's 20 failures hits a spare with a 14-day lead time. That single event puts 336 hours on the clock. Annual downtime goes to 374 hours, and operational availability drops to 7,626 / 8,000, or 95.3%.
The inherent formula never counted that wait, so it keeps showing 99.5%.
What redundancy changes
I use this example in my preventive maintenance course to teach consequence of failure.
Take a single transformer feeding a whole array of production-critical equipment. The failure mode is winding insulation breakdown from sustained overloading.
When it goes, the transformer is out for several months, the downstream plant stops, and the production losses can be many times the cost of the repair.
Now install a spare: 2 transformers, each rated at 100% capacity.
This time the consequence is the damage to transformer 1 and the cost to repair it, with little or no production loss.
Assume each unit on its own runs at 98% availability:
| Configuration | System availability |
|---|---|
| Single unit | 98.0% |
| 2 × 100% (either one is enough) | 1 − (1 − 0.98)² = 99.96% |
| 2 × 50% (both must run) | 0.98² = 96.04% |
That's 2 configurations, one almost 2 points up and one almost 2 points down, with identical transformers.
Redundancy changed what a transformer failure costs you. Each machine is exactly as reliable as it was. The system's reliability and availability went up.
2 cautions before you use that arithmetic on your own plant.
Treat the percentages as illustrative. For parallel arrangements the arithmetic is an approximation. And no standard defines "system availability". ISO 20815 (2018 edition) says a case-specific definition is needed, so write yours down.
The bigger caution is that 99.96% is optimistic. Your 2 transformers probably share a design, an installation crew, a switchroom, a maintenance crew and a procedure.
That's common cause failure. Assume just 5% of failures hit both units at once. The independent term in the unavailability is about (0.95 × 0.02)² ≈ 0.00036. The common cause term is 0.05 × 0.02 = 0.001. That second term is nearly 3 times larger, so system availability lands nearer 99.86%.
If one of the 2 sits on standby, its failure is hidden. It looks fine until you need it, so it needs a failure-finding task.
The power industry worked this out decades ago. Alongside its Availability Factor, the North American Electric Reliability Corporation (NERC) publishes Starting Reliability: actual unit starts divided by attempted starts. That separate metric exists because Availability Factor can't tell you whether the unit will start.
The gap between the 2 numbers is your maintenance organisation
The gap between inherent and operational availability belongs to your maintenance organisation. It covers your planning quality, your scheduling discipline, whether the spare was on the rack or on a boat, whether the permit was ready at 7am or 2pm, and whether a job waited 3 weeks for a shutdown window it could have shared with 4 others.
Report inherent and operational availability side by side, and the difference is your maintenance organisation's score.
Preventive maintenance can only ever achieve the inherent reliability of your plant: the reliability that was designed and built into it. However much PM you do, there's a limit, and it's often well below what the plant was designed for.
Reliability erodes in 2 steps. From specified to inherent, through design compromises, equipment and vendor choices, fabrication defects, poor preservation and sloppy construction. Then from inherent to operating, when the plant starts up into a reactive environment.
You can win most of the second step back with better operations and maintenance practices and defect elimination. Some of the first step you can recover through engineering changes. The rest you live with.
Terrence O'Hanlon put it best: you can't maintain your way to reliability. I've made that case at length in a separate article.
Is there such a thing as "world class availability"?
Not one you can rely on. SMRP's Metric 2.2 has a best-in-class field, and it reads:
"SMRP's Best Practices Committee research indicates that best-in-class values for this metric are highly variable by industry vertical and facility type. SMRP recommends organizations become involved in trade associations within their industry vertical, as these groups often publish such data about their industry."
The leading maintenance and reliability professional body researched a best-in-class availability value and chose not to publish one. The same metric adds: "Do not confuse availability with reliability."
The 90% and 95% figures you'll see quoted most likely trace back to Nakajima's work on total productive maintenance. It set ideal equipment targets of 90% availability, 95% performance and 99% quality, for an overall equipment effectiveness (OEE) of about 85%. Those were equipment targets for discrete manufacturing, never a plant availability benchmark. My suspicion is that the 95% availability figure is the performance number read off the wrong line, but I can't prove it.
Maintenance cost benchmarks have the same problem.
The widely published guidance that maintenance spend should sit between 2% and 5% of replacement asset value can do serious damage in a capital-intensive business. I once ran that calculation for an LNG plant. Our budget came in at around 0.3% of replacement asset value, less than a sixth of even the 2% low end, and I'd already set a target to take out about 20% of our costs.
Had I used the published benchmark, I'd have been way off in the wrong direction. A maintenance manager from a UK company once called me with exactly the same problem. He'd read that 2% was a good target and couldn't work out why his budget was a fraction of it.
What real availability data does exist?
Power generation is the one sector that publishes real availability statistics.
NERC's Generating Availability Data System covers North American power generation and is free to download. From the 2021 to 2025 brochure, all units reporting:
| Category | Units | Availability Factor | Equivalent Forced Outage Rate |
|---|---|---|---|
| Nuclear | 97 | 92.38% | 2.22% |
| Fossil | 908 | 78.74% | 15.15% |
| Combined cycle | 354 | 85.94% | 6.22% |
| Gas turbine | 549 | 88.59% | 44.44% |
| Hydro | 899 | 78.17% | 10.77% |
In the gas turbine row, 88.59% availability against an equivalent forced outage rate of 44.44%. Peaking units run few hours, and that outage rate divides by service hours, so it overstates the problem. But the point stands. Reserve shutdown counts as available, whether or not the unit would have started.
The International Atomic Energy Agency's PRIS database covers nuclear power worldwide. The energy availability factor for 2023 to 2025 was 83.5% across 411 reactors, with the USA at 92.9% and France at 69.3%. Those figures are energy weighted, so a unit derated to half power all year scores roughly 50%.
Both are measured population statistics you can compare yourself against.
Outside power generation there's very little. Solomon benchmarks refinery availability, but only shares it with participants. EN 15341 defines the indicators and sets no targets.
No credible body I know of publishes a plant availability benchmark. The number depends too much on which clock you started.
How does an availability target get gamed?
Mostly by using the published definitions exactly as written.
NERC's own equations are the cleanest example. Available hours are service hours plus reserve shutdown hours, plus a couple of smaller categories. Availability Factor is available hours divided by period hours.
A unit that spends a whole month shut down in reserve, generating nothing, scores 100% Availability Factor.
There are 2 other common ways, both within the rules. Measure against scheduled time, and an asset nobody scheduled can't be unavailable. Exclude events "outside management control", and the hard ones disappear. A North American grid operator has refused one such exclusion on the record, because it would reward a generator for staying broken longer.
What should you do with your availability number?
4 things, and none of them needs new software.
1. Write down your clock. Put the definition in the metric's own documentation: which hours are in the denominator, what counts as down, where preventive downtime goes, and whether logistic delay is in or out. Can anyone at your site answer that in one sentence? If not, your number isn't under control.
2. Report inherent and operational availability side by side. One is the design's promise and the other is your delivery. The gap between them shows you where to put your improvement work.
3. Pair availability with a reliability measure. Availability alone can't tell you whether an asset will run a full campaign. Add the probability of completing a defined interval, a failure count, or a starting reliability figure for anything that sits idle waiting to be called on.
4. Be careful about copying a target. I set hard targets where they're justified. I expect PM compliance above 95%, ideally above 98%, month in month out. That target works because I control the definition, and because on-time PMs are what let you judge whether your PMs work. For the opposite reason, I warn people off best-practice targets for the PM/CM ratio: other plants define it differently.
Only set a hard target on a metric whose definition you control. Availability definitions vary between plants, so don't copy someone else's availability target.
Where this leaves you
Spares on the rack, a bigger crew, redundancy and a faster permit process all push availability up without making a single machine more reliable. Reliability comes from design, operating practice and defect elimination.
So the next time somebody puts an availability percentage in front of you, ask 2 questions. Which hours were in the denominator? And what happened to the delays?
If they can't answer both, fix the definition before you try to manage the number.



