October 17, 2026
Posted by: editor
Category: RECD & Emission Control Devices

“Our RECD achieves 74% ± 5% particulate reduction.”

It sounds precise and reassuring, but the sentence leaves several questions open. Is that ±5 percent of 74, or ±5 percentage points? How confident is the “±”? Was it one run or ten? And does a result of 74% ± 5% definitely pass a 70% requirement, or only probably?

Every emission measurement carries uncertainty, and an RECD reduction claim combines several measurements, so its uncertainty is larger than any single reading. This guide explains how to read the uncertainty in an RECD emission reduction claim, where it comes from, how to calculate it for a reduction percentage, and how to decide whether a result really clears the 70% line.

Start by Identifying What Kind of Claim You’re Reading

Reduction claims come in several forms, and each tells you something different:

Claim as written	What it tells you	What it leaves out
“Meets the 70% requirement”	A pass/fail statement	The actual result, the margin and the uncertainty
“74% reduction”	A single best estimate	How reliable it is
“74% ± 5%”	A best estimate with a range	Whether ± is percent or points; confidence level
“74% ± 5 percentage points (k = 2, about 95% confidence)”	Estimate, range and confidence	Fully specified
“74% (95% confidence interval 69–79%)”	Estimate and range from repeated runs	Fully specified
“Up to 90% reduction”	The best result observed under some conditions	Typical performance and the conditions behind the figure

The more fully specified a claim is, the more you can do with it. “Up to” figures describe the top of a range, not what you should expect on your engine. Ask for the conditions and the typical result. “Meets the requirement” on its own can’t be checked at all.

The Vocabulary You Need

Uncertainty versus error

Error is the difference between a measured value and the true value, which you never know exactly. Uncertainty is a quantified estimate of how far the true value might reasonably be from the measured one. A good lab reports uncertainty because it can’t report error.

Standard and expanded uncertainty

A standard uncertainty is expressed as one standard deviation. An expanded uncertainty multiplies it by a coverage factor (k), typically k = 2. That gives a range expected to contain the true value with about 95% confidence. Labs accredited to ISO/IEC 17025 usually report expanded uncertainty. The internationally used method is set out in the Guide to the Expression of Uncertainty in Measurement (GUM).

If a report gives “±” without stating k or a confidence level, ask.

Percent versus percentage points

This causes more confusion than anything else in RECD claims:

74% ± 5 percentage points means the range is 69% to 79%.
74% ± 5% (relative) means ±5% of 74, which is ±3.7 points, giving a range of 70.3% to 77.7%.

The difference is small here but can matter right at the 70% line. Reduction claims should always be in percentage points, and the report should say so.

Random and systematic components

Random uncertainty varies from run to run: sampling scatter, small load fluctuations, weighing noise. Repeat runs reduce its effect on an average.
Systematic uncertainty shifts every result the same way: a calibration bias, a method difference, a basis error. Repeat runs don’t reduce it.

Where the Uncertainty in an RECD Result Comes From

Each concentration measurement in an RECD test has its own uncertainty budget. Typical contributors for a gravimetric particulate measurement:

Source	Type	Comment
Filter weighing (balance, blanks, conditioning)	Random and systematic	Large for small post-RECD filter masses
Sample volume (dry gas meter calibration, leaks)	Mostly systematic	See instrument calibration records
Isokinetic sampling deviation	Random and systematic	Larger at disturbed ports
Temperature and pressure normalisation	Systematic	Small if instruments are calibrated
O₂ or CO₂ correction	Systematic	Grows at light load, where O₂ is high
Representativeness of the sampling point	Systematic	See exhaust sampling location
Process variability (load drift, regeneration)	Random	See load matching

For stack particulate measurements, combined expanded uncertainties of roughly 10–20% of the measured value are common, and they can be larger at low concentrations.

How Uncertainty Carries Into a Reduction Percentage

The reduction is calculated from two measurements:

R = 1 − (C_after ÷ C_before)

The uncertainty of R depends on the uncertainties of both concentrations. If the two measurements’ errors are independent, a common approximation is:

U(R) ≈ (C_after ÷ C_before) × √[ (U_after ÷ C_after)² + (U_before ÷ C_before)² ]

Here U is the expanded uncertainty and U ÷ C is the relative expanded uncertainty.

Worked example

The figures are illustrative.

Before: C_before = 80 mg/Nm³, expanded uncertainty ±10% (relative)
After: C_after = 22 mg/Nm³, expanded uncertainty ±12% (relative). It’s larger because less mass was collected.

Step 1, best estimate:

R = 1 − (22 ÷ 80) = 1 − 0.275 = 72.5%

Step 2, relative uncertainty of the ratio:

√(0.12² + 0.10²) = √(0.0144 + 0.0100) = √0.0244 = 0.156

Step 3, absolute uncertainty of R:

U(R) = 0.275 × 0.156 = 0.043, or ±4.3 percentage points

Result: 72.5% ± 4.3 percentage points, a range of about 68.2% to 76.8% at roughly 95% confidence.

Notice that the uncertainty of R is driven by the ratio (0.275), not by the reduction itself. The higher the reduction, the smaller the ratio, and the narrower the absolute uncertainty on R. A device achieving 90% has a ratio of 0.10, so the same relative uncertainties would give only about ±1.6 points.

A note on correlated errors

Some systematic errors affect both measurements the same way, for example a slightly biased barometer used for both. These partly cancel in the ratio. The formula above assumes independent errors, so it tends to overestimate the uncertainty when the same instruments and method are used for both readings. That makes it a reasonable, slightly cautious estimate. For a same-day upstream/downstream test with identical equipment, the true uncertainty of R may be a little smaller.

Repeat Runs: Measuring the Random Part Directly

When a test includes several runs at the same load, the spread of results shows the random part of the uncertainty directly.

Worked example with three runs

The figures are illustrative. Reductions measured in three valid runs: 71.0%, 74.5%, 72.6%.

Mean: (71.0 + 74.5 + 72.6) ÷ 3 = 72.7%
Standard deviation: 1.75 points
Standard error of the mean: 1.75 ÷ √3 = 1.01 points
95% confidence interval: with only three runs, use the t-distribution factor for 2 degrees of freedom, 4.30. That gives 72.7 ± 4.30 × 1.01 = 72.7 ± 4.3 points, or about 68.4% to 77.0%

Even though the three runs are within 3.5 points of each other, the 95% interval for the mean is wide. That’s because three runs give limited information about the true spread. With six runs of similar scatter, the interval would narrow to roughly ±1.8 points.

This interval covers only the random part. Systematic uncertainties, such as calibration, method and basis, sit on top of it.

When You Only Have One Run

Many site tests, especially older ones or those done for consent monitoring, include only one run at each load point. A single run gives no direct information about run-to-run scatter, so its uncertainty has to be estimated entirely from the uncertainty budget.

That is a weaker basis:

The random component is estimated, not observed. The lab relies on typical values for sampling and weighing variability, which may not reflect your site’s conditions.
Outliers can’t be spotted. If something unusual happened during the run, such as a load swing or a brief regeneration, there is nothing to compare it with.
Confidence in a borderline result is low. A single run at 71% tells you little about whether the device reliably exceeds 70%.

If you hold a single-run result:

Check the run’s field records for load stability, thermal stability and regeneration status.
Look at the stated uncertainty and ask how it was estimated.
If the result is near 70%, or if a decision depends on it, plan a repeat test with at least three runs per load point.

Judging a Result Against the 70% Line

Once you have a best estimate and an uncertainty range, there are four possible positions relative to a 70% requirement:

Zone	Example result	Range	Interpretation
Clear pass	78% ± 4	74–82%	Whole range above 70%
Pass, but range crosses the line	72% ± 4	68–76%	Best estimate passes, but compliance isn’t certain
Fail, but range crosses the line	68% ± 4	64–72%	Best estimate fails, but compliance can’t be ruled out
Clear fail	63% ± 4	59–67%	Whole range below 70%

The two middle zones are where disputes happen. How to handle them is a matter of policy, not measurement. That policy is called a decision rule.

Common decision rules

International guidance for laboratories, such as ILAC-G8 on decision rules and statements of conformity, describes several approaches. Applied to RECD testing:

Simple acceptance (shared risk): pass if the best estimate is at or above 70%. Easy to apply, and the risk of a wrong decision is shared between buyer and supplier.
Guarded acceptance: pass only if the lower end of the uncertainty range is at or above 70%. This protects the buyer, but a good device can fail.
Guarded rejection: fail only if the upper end of the range is below 70%. This protects the supplier, but a weak device can pass.

None of these is “correct.” What matters is choosing one before testing and writing it into the contract or test protocol. Choosing a rule after seeing the result is how a measurement turns into an argument.

Applying the rules to the worked example

For 72.5% ± 4.3 points (range 68.2–76.8%):

Simple acceptance: Pass (72.5 ≥ 70)
Guarded acceptance: Not a pass (68.2 < 70). Retest, or run more repeats to narrow the range.
Guarded rejection: Not a fail (76.8 ≥ 70)

The same data gives three different contractual outcomes depending on the rule.

Uncertainty in Weighted Multi-Load Results

Some reports give a single reduction figure that combines several load points, weighted by how much time the engine spends at each. Type-approval tests use a weighted multi-mode cycle such as ISO 8178 D2, and some site tests borrow the approach.

A weighted figure hides how uncertainty varies between modes:

Light-load modes usually carry the largest relative uncertainty, because exhaust is cooler, concentrations after the RECD are low, and O₂-based corrections are more sensitive.
High-load modes usually have smaller relative uncertainty, because there is more particulate to collect and conditions are more stable.

When you review a weighted result:

Ask for the reduction and its uncertainty at each mode, not just the weighted total.
Check whether any single mode falls below 70% within its uncertainty range, even if the weighted total doesn’t.
If your set spends most of its time at light load, give that mode’s result particular attention. See choosing an RECD for a generator running at low load.

Is a Change Between Two Tests Real?

Uncertainty also matters when comparing results over time. Suppose the acceptance test gave 76% ± 4.0 points, and a periodic test a year later gave 72% ± 4.3 points. Has the device deteriorated?

A simple check compares the difference with the combined uncertainty of the two results:

Difference = 76 − 72 = 4.0 points Combined uncertainty = √(4.0² + 4.3²) = √(16.0 + 18.5) = √34.5 ≈ 5.9 points

The difference (4.0) is smaller than the combined uncertainty (5.9), so the two results are consistent with each other. The data doesn’t show that the device has deteriorated. It also doesn’t prove that it hasn’t.

If the second result had been 66% ± 4.3, the difference would have been 10 points, well beyond 5.9. That would be good evidence of a real change, worth investigating. See can RECD efficiency change between new and aged conditions.

This only works if both tests used comparable conditions and methods. Otherwise, part of the difference may come from load, basis or engine changes, not the device. See how to compare RECD test results from different laboratories.

Questions to Ask the Lab About Uncertainty

Is the stated uncertainty expanded, and what coverage factor or confidence level was used?
Is it given for each concentration and for the calculated reduction?
Is it expressed in percentage points for the reduction?
Which sources were included in the uncertainty budget?
Does it include run-to-run variability, and how many runs were made?
How were correlated errors between the before and after measurements handled?
Is the uncertainty estimate part of the lab’s accredited scope?

A lab that can answer these clearly has probably done the work properly.

Sample Contract Wording for a Decision Rule

“The RECD shall be accepted if the mean particulate reduction at each specified load point, calculated from at least three valid runs, is not less than 70%. The test report shall state the expanded uncertainty of the mean reduction (k = 2) in percentage points. Where the mean is not less than 70% but the lower bound of the expanded uncertainty range is below 70%, the parties shall [accept the result / conduct up to three additional runs and recalculate / …].”

Adapt the bracketed choice to your risk appetite. Our guides on writing measurable performance requirements into RECD tenders and RECD performance guarantees and operating conditions explain where this clause fits.

Ways to Reduce Uncertainty Before the Test

If you expect a result near 70%, reducing uncertainty in advance is cheaper than arguing afterwards:

Run more repeats at each load point. This shrinks the random component.
Sample longer after the RECD, so more mass is collected and weighing uncertainty becomes relatively smaller.
Use the same instruments and method for the before and after measurements, so systematic errors partly cancel.
Prefer a same-day upstream/downstream design, which removes engine, fuel and ambient variability. See how fuel and engine condition can confound RECD performance tests.
Control load, warm-up and regeneration tightly. See how to account for warm-up time and RECD testing during regeneration.
Insist on complete calibration records. See which instrument calibration records should accompany RECD tests.

Red Flags in Uncertainty Statements

No uncertainty at all. The result can’t be judged against a threshold objectively.
“±” with no unit or confidence level. Ask whether it’s percent or percentage points, and at what k.
Uncertainty for the concentrations but not for the reduction. The reduction’s uncertainty must be calculated, and it is usually larger than people expect.
Implausibly small uncertainty, such as ±0.5 points on a single gravimetric run. Ask how it was estimated.
A decision rule chosen after the result was known.
Only the best run reported, with others omitted.

Explaining Uncertainty to Non-Technical Readers

Boards, residents and auditors don’t need the formulas. They need an honest summary, for example:

“Our best estimate is that the device removes about 72% of particulate by mass. Allowing for measurement uncertainty, the true figure is likely between 68% and 77%. Our contract specifies that a further test will be carried out to confirm compliance with the 70% requirement.”

That is more credible than a bare “72%”, and far more credible than “up to 90%”.

Frequently Asked Questions

Does 72% ± 4% mean my RECD passes the 70% requirement?

It depends on the decision rule. Under simple acceptance, the best estimate of 72% passes. Under guarded acceptance, which requires the whole range to be above 70%, it doesn’t pass, because the range extends down to 68%. Agree the rule before testing.

Why is the uncertainty of the reduction larger than I expected?

Because the reduction combines two uncertain measurements, and the post-RECD measurement often has the larger relative uncertainty, since less particulate is collected. The combined effect is usually a few percentage points.

Will running more tests always improve the result?

More runs narrow the random part of the uncertainty and give a more reliable average. They don’t remove systematic errors, and they won’t change a device’s true performance. They make the estimate more trustworthy, whether it ends up above or below 70%.

What is a coverage factor?

It is the multiplier applied to a standard uncertainty to produce an expanded uncertainty with a stated confidence. k = 2 gives about 95% confidence for normally distributed errors and is the most common choice.

Should the supplier or the buyer bear the risk of measurement uncertainty?

That’s a commercial decision. Simple acceptance shares the risk, guarded acceptance shifts it to the supplier, and guarded rejection shifts it to the buyer. The important thing is to decide in advance and write it down.

Treat the Uncertainty as Part of the Result

A reduction figure without its uncertainty is half a result. Aceget supplies CPCB type-approved Retrofit Emission Control Devices for DG sets from 25 kVA to 2500 kVA. We encourage clients to agree decision rules and uncertainty reporting before acceptance testing. Choose a lab accredited by NABL for stack emission monitoring, review our certifications, and contact our team for help interpreting a result.

compliance confidence interval decision rules emission reduction measurement uncertainty RECD testing

How to Interpret Uncertainty in an RECD Emission Reduction Claim