Help reading DMM Calibration report

[jeffjmr]

I found the definition of TAR as Test Accuracy Ratio, but can’t correlate that with the results of this report. Could someone educate me?

Thanks,
Jeff

[Kleinstein]

The TAR value is the allowed error of the instrument to test relative to the uncertainty of the instrument / reference used for the calibration. This is another way to express how good or bad the calibrator is.
So the TAR numbers say nothing about the meter tested. It is about the quality (or at least their level of confidence/certification) of the cal lab relative to the requirements for that test.

The 999.9 number looks like a way to say > 999 - which is easy for the short in the Mohms range.

With a TAR of >10 there is usually no problem with the reference and no extra uncertainty to add to the calibration. A TAR < 5 may need extra considerations / explaination (e.g. use 2 independet refs, extra guard instrument). This may lead to high uncertainty in the calibration.

[alm]

To expand on that, TAR is calculated as:
$$TAR=\frac{\textrm{Tolerance of device under test}}{\textrm{Accuracy of calibration source}}$$

So to calculate it you need the specified tolerance of your DMM at whatever calibration interval it's calibrated to (generally 1 year unless the customer requested otherwise) and the accuracy of the calibrator that's being used. If you know they use a Fluke 5700A that is calibrated every 90 days, you could look up the specs and get an idea what the accuracy might be.

[mendip_discovery]

To expand on that, TAR is calculated as:
$$TAR=\frac{\textrm{Tolerance of device under test}}{\textrm{Accuracy of calibration source}}$$

So to calculate it you need the specified tolerance of your DMM at whatever calibration interval it's calibrated to (generally 1 year unless the customer requested otherwise) and the accuracy of the calibrator that's being used. If you know they use a Fluke 5700A that is calibrated every 90 days, you could look up the specs and get an idea what the accuracy might be.

Would it not be possible to just calculate the accuracy by dividing the spec applied by the TAR

Range	Applied	Lower	Upper	Measured	TAR	Spec	Accuracy
30 Ω	9.9995	9.992	10.007	9.9981	32.60	0.015	0.00046
300 Ω	99.991	99.969	100.013	99.986	15.50	0.044	0.00284
3 kΩ	0.99992	0.99974	1.0001	0.99987	18.00	0.00036	0.00002
30 kΩ	9.9997	9.9979	10.0015	9.9992	18.90	0.0036	0.00019
300 kΩ	99.994	99.976	100.012	99.989	16.40	0.036	0.00220
3 MΩ	0.99991	0.99973	1.00009	0.99988	12.00	0.00036	0.00003
30 MΩ	9.9984	9.9904	10.0064	10.0004	25.80	0.016	0.00062

I have never been much of a fan of the TAR/TUR system but that is partly as it is just an older way of thinking and I just don't use it. I feel for a calibration certificate they should be using TUR as that is more useful to the customer. The Spec of my meter is better than yours, ignoring all the other factors that can affect the results.

[guenthert]

To expand on that, TAR is calculated as:
$$TAR=\frac{\textrm{Tolerance of device under test}}{\textrm{Accuracy of calibration source}}$$

So to calculate it you need the specified tolerance of your DMM at whatever calibration interval it's calibrated to (generally 1 year unless the customer requested otherwise) and the accuracy of the calibrator that's being used. If you know they use a Fluke 5700A that is calibrated every 90 days, you could look up the specs and get an idea what the accuracy might be.

      Hmmh.  That might be the definition, but that lab clearly computed tar as
$$TAR=\frac{\textrm{specified tolerance}}{\textrm{deviation from expected value assuming perfectly known artefact}}$$

Which seems to be quite questionable practice.

[bdunham7]

A good explanation of the difference between TAR and TUR from a reputable but non-electronics related source:

https://www.mitutoyo.com/webfoo/wp-content/uploads/15005A.pdf

I'd like to see that whole calibration report, including the lab information.  Perhaps they can be asked to comment and perhaps justify their method of calculation. 

[bdunham7]

      Hmmh.  That might be the definition, but that lab clearly computed tar as
$$TAR=\frac{\textrm{specified tolerance}}{\textrm{deviation from expected value assuming perfectly known artefact}}$$

Which seems to be quite questionable practice.

I'm not seeing that as matching their results.  All of the zero lines in the first batch should be 999.9  (0/0 or overflow or infinity or ...) by your formula, but they're not. 

[mendip_discovery]

Trying to divide a number by zero can give some odd numbers. Wondered if they had tried to use some sort of uncertainty in the zero checks. I think they that if this was a 17025 cert then the calculations for the zero checks wouldn't be allowed.

[bdunham7]

Trying to divide a number by zero can give some odd numbers. Wondered if they had tried to use some sort of uncertainty in the zero checks. I think they that if this was a 17025 cert then the calculations for the zero checks wouldn't be allowed.

I don't think they are dividing by zero.  I think the first batch of results--the zero points--are simply telling us that the test short has an 'accuracy' of +/- 0.00005R, or 50 microohms.  The second batch is using a set of standard resistors, perhaps in a calibrator, that each have their own individual tolerances that could be reverse-calculated using your method, although I didn't check the math.  For the 30M example, it appears that the DUT is specified as +/- 8k and the test resistor is 9.99840M +/- 0.31k.

[guenthert]

      Hmmh.  That might be the definition, but that lab clearly computed tar as
$$TAR=\frac{\textrm{specified tolerance}}{\textrm{deviation from expected value assuming perfectly known artefact}}$$

Which seems to be quite questionable practice.

I'm not seeing that as matching their results.  All of the zero lines in the first batch should be 999.9  (0/0 or overflow or infinity or ...) by your formula, but they're not.

Neither do I.  I'm having my coffee now.  It bugs me though, that there's no uncertainty given for the 'applied' value.

[bdunham7]

Neither do I.  I'm having my coffee now.  It bugs me though, that there's no uncertainty given for the 'applied' value.

Somewhere in the calibration report there should be a list of equipment used and the specifications/uncertainty assigned to each.

[mendip_discovery]

They maybe using a 8.5digit meter as the reference and using the short term spec of that on the resistor to get quite good numbers for the TAR. Using a spec rather than a Uncertainty they are satisfying a cursory read of the cert show a good ratio.

The 30 ohm range is the one that interests me but not looked at the numbers properly yet.

[Kleinstein]

The reference resistor accurcy seems to be at around 2x the specs for the Fluke 5700A. So good (definitely good enough for a 5.5 digit meter), but not extra-ordinary.

[jeffjmr]

You have all touched on the reason for my initial inquiry.

Here’s the equipment used for the cal. It has been almost 10 months since the calibration.

It’s for my HP3478A by the way in case that helps.

Thanks all,
Jeff

[alm]

For the 10 kOhm point, the uncertainty of the 3478A with a 1 year calibration interval is:
$$160 \textrm{ppm} + \frac{2 \cdot 100 \textrm{mOhm}}{30 \textrm{kOhm}} = 166.67 \textrm{ppm}$$
So based on the TAR of 18.9 this would give the calibrator an uncertainty of 8.8 ppm. Given that the 5700A is about 4 months away from calibration, it's clearly on a 180 days or 1 year calibration cycle. The quoted uncertainty for the 10 kOhm range is 10.5 ppm (180 days) and 12 ppm (1 year). So I'm not sure how they arrived at that TAR figure.

[bdunham7]

So I'm not sure how they arrived at that TAR figure.

Using the 90-day 95% figure of 9.5 for the calibrator and 180ppm for the 3478A (derived from the limits shown on the cal certificate) gives you 18.9.   It doesn't matter what calibration cycle it is on, just how long since the last calibration.

[alm]

It doesn't matter what calibration cycle it is on, just how long since the last calibration.

Doesn't it? Doesn't a standard calibration involve a performance verification against the limits determined by the calibration interval, and if the unit is inside these limits plus some guard band, it's declared in spec and gets a calibration certificate? So just because it was calibrated 24h ago does not mean its inside the 24h specs. Now in the case of the 5700A it might be that they always do an artifact calibration and that the artifact standards used are good enough to meet the 90 day specs (which are very close to the 180 day and 1 year specs). But that's not obvious to me from the information that they provide.

[bdunham7]

Doesn't it? Doesn't a standard calibration involve a performance verification against the limits determined by the calibration interval, and if the unit is inside these limits plus some guard band, it's declared in spec and gets a calibration certificate? So just because it was calibrated 24h ago does not mean its inside the 24h specs. Now in the case of the 5700A it might be that they always do an artifact calibration and that the artifact standards used are good enough to meet the 90 day specs (which are very close to the 180 day and 1 year specs). But that's not obvious to me from the information that they provide.

Good point, you always need to read the actual certificate I suppose.  IMO, if you don't calibrate to the 24h specs, the unit can't be assured to meet the manufacturer specs even for a 1-year calibration interval.  Part of that is based on the fact that for many devices, there are 24h/1C specs and then 90day/5C specs.  My take is that unless you calibrate the device to the 24h/1C specs, you have done a second-rate job at best.  But that's a complex issue and it isn't the way things are always done.  And then there's the 'relative to standards' issue.

[Kleinstein]

Some of the meassurement error is due to noise / INL of the DUT. To take this fully into account the test limits for the calibration should be different (stricter) than the specs for the meter. This also leads to borderline cases when the tests can not for sure say that the DUT is bad or good. A low TAR makes this unclear zone wider, but even with best TAR this case can still happen.

[mendip_discovery]

What the heck does 24hr specs got to do with an annual calibration? It's how well that device performs over a 24hr period. So if you were to make a measurement at 8am and another at 8pm then there is a good chance your measurments won't be outside that specification.

You would only adjust the UUT if it is at a agreed percentage of the annual spec, or even the customers spec/requirements. The 24hr spec would be pointless spec to use for that.

The lab may have history for thier devices and so are working to thier own specifications. The likes of fluke etc will give the typical specs, many of the meters are better than that in practice. Though there will plenty of companies out there with specs that are just dreamy up to sell the device. Also any 17025 lab will also be doing regular cross checks on the equipment to build confidence that the test gear they have is behaving and not drifting off.

[alm]

So I guess the conclusion is that the accuracy of the calibrator might be a sensible figure, but that the tolerances used for the 3478A are too wide to confirm it's within its 1 year manufacturers specifications.

[bdunham7]

So I guess the conclusion is that the accuracy of the calibrator might be a sensible figure, but that the tolerances used for the 3478A are too wide to confirm it's within its 1 year/5C manufacturers specifications.

I'm not seeing that and looking again I think you miscalculated the tolerances.  The HP manual lists the 30k range as .016% of reading plus 2 counts.  That works out to 18 counts at 10k, which is what they used.  In any case, if you get a fully documented calibration with data, then you don't have to wonder so much--you can see for yourself.  For example, the 24h/1C specs for the 30k range amount to 6 counts and the deviation of the meter from the reference is 5 counts, so it is within the 24h/1C specs relative to the calibration source.

[bdunham7]

You would only adjust the UUT if it is at a agreed percentage of the annual spec, or even the customers spec/requirements. The 24hr spec would be pointless spec to use for that.

And this is why a calibration certificate in general means so little to me.  This may be a widespread practice, but it does not mean that the meter meets the original manufacturer's specifications and confidence interval over the stated time and temperature spans.   Calibrating the instrument to the 24h/1C specs in a +/- 1C environment does give that assurance.  And there are calibration labs that do it that way by default.

[bdunham7]

You have all touched on the reason for my initial inquiry.

Here’s the equipment used for the cal. It has been almost 10 months since the calibration.

It’s for my HP3478A by the way in case that helps.

Did you mean that your meter was calibrated 10 months ago or that the labs calibrator was calibrated 10 months ago?

All in all, it actually looks like they've done a good job, IMO, even if there are questions or small errors in the TAR calculation.  A healthy overkill margin is one way to make up for such minor issues.

[alm]

I'm not seeing that and looking again I think you miscalculated the tolerances.  The HP manual lists the 30k range as .016% of reading plus 2 counts.  That works out to 18 counts at 10k, which is what they used.

You're completely right, I reversed the calculation of the uncertainty, converting counts to ppm of range, instead of ppm of value. So the tolerance for the 10k range does match the 1 year specification.