which confidence interval was used on old Fluke specifications?

[e61_phil]

Hi,

does anyone know which confidence intervall (95%?, 99%?) was used for old Fluke specifications as listed in the Fluke 5440B or 5450A manual?

And why did they say one has to add the uncertainty of the cal standard algebraically (old manuals)? In newer specifications (8508A for example) they RSS the cal standard uncertainty together with the measurement uncertainty.

Best
Philipp

[CalMachine]

I would bet it would be 99%, since their certificates reflect a 99% confidence interval.

[Fsck]

I'd place my bets on 99% as well, as that's their metrology standard. (I'd bet on 95% for "consumer" meters)
I remember reading something somewhere about it, but my memory is foggy, thus only betting.

[CalMachine]

What I really don't like is mostly all measurement uncertainties are reported @ a 95% confidence interval, while the DUT performance is verified at a 99% confidence interval.     

[AG7CK]

Hi,

does anyone know which confidence intervall (95%?, 99%?) was used for old Fluke specifications as listed in the Fluke 5440B or 5450A manual?

And why did they say one has to add the uncertainty of the cal standard algebraically (old manuals)? In newer specifications (8508A for example) they RSS the cal standard uncertainty together with the measurement uncertainty.

Best
Philipp

Even if I don't know, I wouldn't bother to check. Surely it is 99%. No sane person would warrant a multi-function calibrator at 95%. Here is the reason (IMO good reading):

" ...
For a standard resistor, whose value is specified at a
95% confidence level to be within limits centered
about the nominal value for the resistor, we
could expect to have to re-trim 5% of the
standards before they could be shipped.
Specifying a complex instrument at the true 95%
confidence level for each point would be a
manufacturing disaster. For example, each
Fluke Model 5520A Multiproduct Calibrator is
tested at 552 points on the production line prior
to shipment. If each of the points has a 95%
probability of being found in tolerance, there
would only be a 0.95552(raised to the 552) = 0.00000000051%
chance of finding all the points within the
specification limits if the points are independent!
Even if we estimate 100 independent points
(about 2 per range for each function), we would
still have only a 0.95100(raised to the 100) = 0.6% chance of being
able to ship the product.
...
"
[Edit: Added two exponents lost in copy from pdf-file]
Source: download.flukecal.com/pub/literature/msc04.pdf
HAVING CONFIDENCE IN SPECIFICATIONS
David Deaver
Fluke Corporation
PO Box 9090
Everett, WA 98206
425-446-6434
David.Deaver@Fluke.com

-o-

Of course it is a bit different if not all points are independent, the box has ACAL or ... But the point is clear: The more output/measurement types/ranges, the more likely it is that one or more test points break the interval. So you want a "wide" interval to at least be sure most of your boxes are within spec for the calibration period.

[e61_phil]

Thanks for your replies!

I also asked Fluke and they responded very quick: "Yes you are correct the 5440B/5450A specs are 99% (k = 2.56) specs."

@AG7CK: I'm not sure if the calculation with 95% probability of failure is really correct. If this would really be the case every calibrated instrument with more than 20 measured points will fail in some points.

[dacman]

Adding the specifications accounts for a worst-case scenario, and used to be done a lot prior to NVLAP.  RSSing them is statistically correct if they are not correlated.  Fluke does not guarantee the 95% specifications, but they do guarantee the 99% specifications (of test equipment that they sell).  The 95% measurement uncertainty (which is not a specification) on reports is an NVLAP requirement for a certified calibration (in the USA).

[e61_phil]

Adding the specifications accounts for a worst-case scenario, and used to be done a lot prior to NVLAP.  RSSing them is statistically correct if they are not correlated.  Fluke does not guarantee the 95% specifications, but they do guarantee the 99% specifications (of test equipment that they sell).  The 95% measurement uncertainty (which is not a specification) on reports is an NVLAP requirement for a certified calibration (in the USA).

Interesting! Does that mean it is right to RSS the spec relative to the cal standard (given in the manual) with the uncertainty of the measurement? (I used a Fluke 8508A with 90day specs). Both should be uncorrelated.

Your point "99% is guranteed and 95% not" ist also interesting. If I assume a normal distribution and Fluke says that the specification of the 5450A/5440B is for k=2.56 isn't it possible to multiply the given specification by 2/2.56 to get the k=2 specification?

Edit: Ok, I tried the 2/2.56 factor for the 99% -> 95% Fluke 5700A specifications and the result is too low
The factor between 99% and 95% of the Fluke 5700A resistance specs seems to be around 0.825

[Moon Winx]

Yeah, if anyone knows or can figure out how Fluke calculated the 95% specs please inform. It's been a mystery around our lab for years.

As a guess, maybe they used a student T distribution due to a small sample size, but I have no idea.

By the way, you can thank the US Air Force for insisting Fluke include 95% specs in their contract buys. It may be useful to them, but it just causes confusion to everyone else.

[dacman]

It should be valid to combine the two, but the spec is relative to calibration standards, which is not the same as measurement uncertainty.  (The calculation would be high.)  For the 5700A relative specifications, the manual cal needs to be done on each range the relative spec is to be used on (and combined with the spec of the calibration standard).

For the 95% and 99% specifications of the 5700A, CIs of 1.96 or 2.576 only apply if the data is Gaussian (as you mentioned), which can be assumed in the absence of other data, according to NIST TN 1297.  But if the data is actually rectangular or U-shaped or triangular (or something else) then the k factors would (usually) be different.  If the Fluke 95% and 99% specifications are based on statistical analysis, then the data would be somewhat triangular.  I would think that the Fluke specifications would also account for equipment reliability and would, therefore, need to be expanded further (and not a specification at the moment of calculation only).

The below was copied from a Fluke 8508A manual:
Notes to Performance Specifications
Fluke guarantees 8508A performance verification using specifications stated to 99 % confidence level.

The below was copied from a Fluke 5700A/5720A manual:
Fluke guarantees performance verification using specifications stated to 99% confidence level.

[e61_phil]

It should be valid to combine the two, but the spec is relative to calibration standards, which is not the same as measurement uncertainty.

Perhaps I used only the wrong terms or I'm completely wrong

What I did so far was using the recently calibrated Fluke 8508A (<90 days, but caled for 1year specs 95%) to measure my Fluke 5450A and tried to calculate the 1year uncertainty of "my" ohms.

I did the following (10k as an example, I did it for all ranges this way):

- I took the 2.2ppm uncertainty from the calibration certificate (I think this is already too high, because it includes already some influence from the 8508A)
- I looked up the 90day 95% rel. cal std. specification for the 8508A (3.5ppm + 0.25ppm of range) which gives 4ppm for 10k

With this information I can measure the resistance of the 5450A to an uncertainty of sqrt( (2.2ppm)^2 + (4ppm)^2) = 4.85ppm

Next step was to have a look in the 5450A manual and add (because they say one have to add it) 10ppm (1year +/-5°C of Tcal) which gives 14.9ppm for "my" 10k uncertainty given by the 5450A over 1year

The 10ppm ist 99% so one can reduce it at least to 90% of the value, I think. The other question is, is it correct to RSS it together with the 4.85ppm? That will give 10.2ppm

Best regards
Philipp

PS: The 5700A was only an example, because they defined 95% and 99% specs.

[dl1640]

The uncertainty reported on your 8508 certificate should include the uncertainty of calibration standard, may be a well maintained 8508. But this uncertainty may or may not include your 8508 stability over time.

The 5700 has abs. uncertainty spec and rel. uncertainty spec, and they both are specified over a time period, that is done by using rich history data, experiments, analysis, i think.

The rel. Uncertainty does not include the cal standard uncertainty at FLUKE, so they would suggest you use abs. Uncertainty in your uncertainty budget, however if you have better cal standard than fluke does, you calculate your own abs. Uncertainty by combine rel. Uncertainty of 5700 and your cal standard uncertainty.

Normally, i think you would consider your 8508 stability if it is as cal standard.

[Moon Winx]

Use 8508A absolute specifications because the uncertainty reported on your 8508A cert doesn't mean much of anything. Unless it is an unusual report, the uncertainty value listed is the calibrating labs' measurement uncertainty and has nothing to do with the performance of the DUT (your instrument).

Having said that, if you have long-term stability data on your meter you could use the reported uncertainty combined with your stability to come up with a usable spec, but this is a lot of work and takes a long time to create.

[e61_phil]

The uncertainty reported on your 8508 certificate should include the uncertainty of calibration standard, may be a well maintained 8508. But this uncertainty may or may not include your 8508 stability over time.

The 8508A was calibrated by an chracterized Fluke 5720A. There is a paper from Fluke how to do it.

I think the uncertainty given in the certificate includes something from the Fluke 8508A, because the 10k uncertainty given in the certificate is different for normal Ohms and LoI ohms (higher). The certificate shows only the characterzed Fluke 5720A, therefore they should have done both measurements (normal and LoI) with the same standard. Why is the uncertainty different?

The 5700 has abs. uncertainty spec and rel. uncertainty spec, and they both are specified over a time period, that is done by using rich history data, experiments, analysis, i think.

The rel. Uncertainty does not include the cal standard uncertainty at FLUKE, so they would suggest you use abs. Uncertainty in your uncertainty budget, however if you have better cal standard than fluke does, you calculate your own abs. Uncertainty by combine rel. Uncertainty of 5700 and your cal standard uncertainty.

Normally, i think you would consider your 8508 stability if it is as cal standard.

The 8508A has also abs. and rel. uncertainty specs. I used the rel. ones together with the uncertainty from the certificate.

Use 8508A absolute specifications because the uncertainty reported on your 8508A cert doesn't mean much of anything. Unless it is an unusual report, the uncertainty value listed is the calibrating labs' measurement uncertainty and has nothing to do with the performance of the DUT (your instrument).

Having said that, if you have long-term stability data on your meter you could use the reported uncertainty combined with your stability to come up with a usable spec, but this is a lot of work and takes a long time to create.

I want to calculate the uncertainty by my own to benefit from the "young" certificate. Otherwise I have to use the 1year absolute uncertainties.

If I know the uncertainty of the standard they used, why shouldn't it be possible to calculate the 90day uncertainty?

[dl1640]

I almost agree with moon winx.
And i would use manufacture spec of 8508 if i have not enough and reliable data.

At least manufacture spec is approved so it has minimum risk. Uncertainty is about risk management.

通过我的 PRA-AL00 上的 Tapatalk发言

[e61_phil]

I almost agree with moon winx.
And i would use manufacture spec of 8508 if i have not enough and reliable data.

At least manufacture spec is approved so it has minimum risk. Uncertainty is about risk management.

通过我的 PRA-AL00 上的 Tapatalk发言

Ok, can you explain at which point you see the risk?

They wrote 2.2ppm uncertainty in the certificate and I added the 4ppm from the 90days rel. to cal. std. to it

Is there any data I should ask for? I already asked Fluke what the "real" 10k uncertainty during the calibration was.

[dl1640]

you combine the uncertainty in the 8508 certificate and the 90day spec, i think that is already risk managed.

But i would be surprised if fluke tell you the real uncertainty of the cal standard.

If you use the calibrated value of 8508 to calibrate uut, and you trust the uncertainty on your 8508 certificate (if you know the cal standard uncertainty is small enough), then probably to use 90day spec alone is just fine.

通过我的 PRA-AL00 上的 Tapatalk发言

[e61_phil]

Thanks, for your explaination! 

I'm not sure if the uncertainty of their cal standards is small enough for a proper 90day calibration, because it is a cal cert. for 1year specifications. Therefore, I tried to use the uncertainty of the standard during calibration together with the relativ uncertainty of the 8508A.

A Fluke person answered already, but he said he will not get the information before monday. We will see..

[Moon Winx]

OK, so I assume you have a Fluke test report and the test point you are interested in is listed with your 8508's measurement of some known source. The known source's value and the measurement uncertainty are also listed, right?

So you can determine the error of your meter at that point to some uncertainty, which you've explained, and if you constantly correct for the offset measured by Fluke I think you could safely use your calculated combined uncertainty for your measurements using the 90-day spec. The important thing is to correct for the measured offset, otherwise you'll need to use the 1-year spec that Fluke tested to.

[e61_phil]

OK, so I assume you have a Fluke test report and the test point you are interested in is listed with your 8508's measurement of some known source. The known source's value and the measurement uncertainty are also listed, right?

Yes, that is right. The calibration certificate showed a 10k measurement with a given uncertainty.

I took that uncertainty added (RSS) the 90day relative spec (8508A) and added (no RSS) the 1year spec (5450A).
It would be nice if the 5450A can also be RSS added.