1:1000 ratio calibration with 3458A or 5440B (high voltage probe)

[e61_phil]

Hi,

I would like to calibrate a high voltage divider. Therefore, I would apply 1000V out of the 5440B and measure the output voltage with a 3458A.

My question is now, which uncertainties can be assumed?

Both instruments (3458A and 5440B) have an AutoCAL feature which refer all the ranges (except the 2,2V and 220mV of the 5440B and I'm only talking about DCV) to their voltage reference. Therefore, I would assume any drift, except the voltage reference itself, will be canceled out. Is this right?

If this is true, one should be able to use the 24h (30day in case of 5440B) specification for ratio only measurements.

In case of the 1:1000 divider with 1000V of input I would assume the following uncertainties:

use of 3458A:

1000V -> 12ppm + 2.6ppm = 14.6ppm
1V -> 1.8ppm
which gives in worst case 16.4ppm


use of 5440B together with 3458A:

1000V -> 2.9 ppm
1V (in 11V Range, because of AutoCal) -> 6.5ppm
that gives 9.4ppm + 0.4ppm for the 1V transfer using the 3458A
This results in less than 10ppm


Is this correct?

Best Philipp

[chuckb]

Some High Voltage Probes are designed to use the 10 Meg Ohm input resistance of the DVM as the final part of the divider. If this is a Metrology grade divider then that does not apply.

[e61_phil]

Some High Voltage Probes are designed to use the 10 Meg Ohm input resistance of the DVM as the final part of the divider. If this is a Metrology grade divider then that does not apply.

This is a self build divider out of many selected Caddock USF resistors (high voltage part) and vishay foil resistors on the low side. It is only used with high impedance inputs.

[e61_phil]

I think (if all the assumptions here are right) one can even lower the uncertainty by using the 10V range of the 3458A to transfer 10V from the 5440B to the 1V output.

1000V 5440B -> 2.9ppm
10V 5440B -> 2ppm
10V to 1V with 3458A -> 0.65 ppm

results in 5.55ppm. RSS even less

[MisterDiodes]

Is your equipment in a current calibration cycle?  If you don't have both pieces in calibration to one standard, then you effectively have two independent voltage references that have drifted apart an unknown amount; Autocal is only going to trim out one single piece of equipment to its own Vref and no effect on the other - so ideally you want make sure you're only getting your ratio measures from -one- piece of equipment and -one- Vref.  For instance you set your voltage source to approximately 1000V, but don't take that "measure" into account as measured on the 5440 voltage source.  Only use the "measured" Vin to your divider as measured on your 3458a for instance, and compare that to divider output on the 3458a.

The other source of error is that for ratio mode, you're taking measures on different ranges (1000v and 1v or whatever) at different points in time on one piece of equipment, so you want to make sure what the noise level is on that 1000V over whatever time frame.  That 5440 used as a power supply could have a drifty Vref inside, and it'll always autocal and look fine compared to itself - you don't know until you measure against something else for absolute value check.

You have the basic idea.  Since you're changing ranges on your 3458a you have to take that into account.  If you can: try the setup a couple of ways and see if what you get makes sense within your measurement uncertainty range.  Another way is to compare your DIY divider to a KVD setup for .0010000 division on the dials, and you probably don't need a 1kV source.

All of this only applies for DC.  If you plan on using your divider for AC then it's probably not going to work like you thought - you'll need some bypass caps on your resistors to keep the thing tuned up for the bandwidth of interest.  That is not trivial for a DIY project and it's much easier to not re-invent the wheel -  acquire a real HV probe if possible.

Keysight has a paper on their website describing ratio mode and how to calculate uncertainty levels for various ratio mode techniques.   

The main thing to consider is what ChuckB pointed out:  Make sure your measuring the divider output only in high impedance mode - otherwise you'll get a false result.

[e61_phil]

Hi MisterDiodes,

thanks for your response. The equipment is in current calibration. But I think that doesn't matter. And the one year calibration cycle wouldn't allow to use the 24h specs for ABSOLUTE values. But I'm not interested in absolute values in this case.

It is clear to me, that the AutoCal has no effect on the other instrument. But I think that doesn't matter.

My first assumption is: The 3458A can ALWAYS transfer 10V to 1V within 0,65ppm. Even the AutoCal doesn't matter here, because it is only a ratio measurement within a single range.

Second assumption is the 5440B bring all the ratios between the ranges into the 30 days spec due to the AutoCal. The 1000V range of the 3458A isn't that great, therefore I use the 5440B for the transfer. The 3458A isn't really needed for this measurement. Any shortterm stable DMM should also work with substitution (measure the divider output and then connect the 5440B and tune the value to be the same as the divider output. Therefore, no good linearity of the DMM is needed). The benefit of the 3458A is the capability to transfer the 10V from the 5440B to 1V.

I want to use at least 1kV because the divider is used at even higher voltages later.

I did already different checks. All of the experiments fit together. My question is more related to: Is this a clean way of uncertainty calculation?

[Kleinstein]

The error calculation and path a measurement sounds somewhat plausible. However even the 24 h specs include some reference / absolute error. So the 1000 V - 10 V ratio from the 5440 might be even better than from it's 30 days specs. There might be more more specific specs on ratios.

On the other side the accuracy specs usually don't include noise (might not be an issue with long enough sampling of averaging enough values).

Another point might be loading of the divider by the input impedance of the meter. At the PPM level this could be an issue even with the high impedance 3458.

It might be a good idea to repeat the measurement with a few different voltages (and a few ACAL cycles on the 5440B). Though not improving much from guarantied specs, it reduces the chance of hitting not so sweep spots. So the average result likely gets better though no tighter specs.

[MisterDiodes]

e61: Maybe you might have a ballpark calculation, but maybe not too clean... be careful you don't convince yourself of some overly low uncertainty estimate.   Have a look at this and you'll find various example for 1/1 and 1/10 ratios on 3458a, and using various methods:

http://www.keysight.com/upload/cmc_upload/All/Calculating_uncertainty_of_a_Ratio_measurement.pdf

The caveat here is that the measure of 1000V and the measure of 1V should be done only on the same instrument for a decent ratio calculation, and I think you're just considering the 5440 as an approximate power supply.  See the note in the examples above about how the ratio calculations apply when crossing ranges.  "Transferring" a voltage measure across to 3458a might work as you describe but remember that comes with it's own level of uncertainty that you must account for.

My other suggestion:  If you're using this for DC, get a resistance measure of both legs of your divider, hopefully with a bridge since you're wanting ppm-level accuracy.  Do you have an accurate Rref to use for comparison?  Your measured resistance ratio should make sense with your measured voltage ratio.  You could also measure each resistor in the string separately but the best way would be to put the whole thing under bias at the expected working voltage - that will give you the best real world conditions.

If you're going to higher voltages over 1kV, be aware that the whole structure changes with corona discharge, self-induced corona oscillation, board leakage and cleanliness, self heating and all of those sorts of considerations - which I'm sure you've thought about.  You don't want to kid yourself that you've got a great divider at 1kV and then think you've got the same performance at 30kV - it won't exactly work that way unfortunately once you get past around 5kV.  Watch out for insulation breakdown and trace spacing in your board (if you have one).  Those kV ranges adds a whole new level of excitement.  Make sure every joint is covered, smooth joints only (no sharp points) and your resistor working voltages have a healthy safety margin.

Glass or ceramic standoffs help as well as extra disk insulators at each resistor - If you go to the long tube style construction those HV dividers can make a nice little oscillator with the self-reactance + corona discharge when you hit the right input voltage at the wrong time.  Always watch your divider output for signs that your DC system has switched over to AC.  It happens.

Have fun!

[e61_phil]

e61: Maybe you might have a ballpark calculation, but maybe not too clean... be careful you don't convince yourself of some overly low uncertainty estimate.   Have a look at this and you'll find various example for 1/1 and 1/10 ratios on 3458a, and using various methods:

http://www.keysight.com/upload/cmc_upload/All/Calculating_uncertainty_of_a_Ratio_measurement.pdf

I know this document and they also say 0.65ppm for the manual 1:10 transfer.

The caveat here is that the measure of 1000V and the measure of 1V should be done only on the same instrument for a decent ratio calculation, and I think you're just considering the 5440 as an approximate power supply. 

I use the same instrument to "measure" the 1000V and the 1V. I put measure in quotemarks, because I use the 5440B for this transfer. I use the 5440B instead of the 3458A, because of its better 1000V specification.

The 3458A is only used to transfer the divider output to the 5440B.

See the note in the examples above about how the ratio calculations apply when crossing ranges.  "Transferring" a voltage measure across to 3458a might work as you describe but remember that comes with it's own level of uncertainty that you must account for.

My other suggestion:  If you're using this for DC, get a resistance measure of both legs of your divider, hopefully with a bridge since you're wanting ppm-level accuracy.  Do you have an accurate Rref to use for comparison?  Your measured resistance ratio should make sense with your measured voltage ratio. 

I think there is no bridge needed because of the high impedant divider input. Another story is the output impedance together with the input impedance of the 3458A, like Kleinstein said. This is a real source of uncertainty. But this is another story.

You could also measure each resistor in the string separately but the best way would be to put the whole thing under bias at the expected working voltage - that will give you the best real world conditions.

This isn't that easy. I can't float the 3458A up to 10kV and the input impedance is also to low to directly measure above the HV resistors.

If you're going to higher voltages over 1kV, be aware that the whole structure changes with corona discharge, self-induced corona oscillation, board leakage and cleanliness, self heating and all of those sorts of considerations - which I'm sure you've thought about.  You don't want to kid yourself that you've got a great divider at 1kV and then think you've got the same performance at 30kV - it won't exactly work that way unfortunately once you get past around 5kV.  Watch out for insulation breakdown and trace spacing in your board (if you have one).  Those kV ranges adds a whole new level of excitement.  Make sure every joint is covered, smooth joints only (no sharp points) and your resistor working voltages have a healthy safety margin.

Glass or ceramic standoffs help as well as extra disk insulators at each resistor - If you go to the long tube style construction those HV dividers can make a nice little oscillator with the self-reactance + corona discharge when you hit the right input voltage at the wrong time.  Always watch your divider output for signs that your DC system has switched over to AC.  It happens.

Have fun!

Thanks

I know there are a lot of uncertainty sources in HV systems. Even self-heating and voltage coefficients will be higher than the 5.5ppm uncertainty. I want to have a clean way to calculate the naked ratio uncertainty. Afterwards I will add the other things.


And also many thanks to Kleinstein. This gives me much more confidence in the measurements

[Dr. Frank]

This whole calibration deals with ratio uncertainty and offsets, not absolute voltage uncertainty.
Therefore, many of the arguments about the real uncertainty of MisterDiodes do not apply here.

The ratio uncertainty of the 5440B (by its internal autocal process) is not specified at all.
It's only a guess from the addendum, where the autocal feature is described in prosa.
The tricky part of this description is, that for an external re-calibration it is sufficient to calibrate the 5440Bs 10V range only, and then do an internal autocal, omitting the ratio measurements by means of an 752A.
So, the guess would be, that autocal brings the instruments ranges relative to each other, back to the 30day specification, within 0.1ppm, as changes in ratios can be detected at that level.

If you substract the basic 10V uncertainty of 1.5ppm, which mainly describes the 30 day reference drift, from both range specifications, you may estimate the
100:1 ratio uncertainty between 10V and 1000V ranges (F.S.) to be around 1ppm. From the 5442A specification, this would even be 0.5ppm.

I also confirmed that with my own Hamon divider, which I calculated to be precise to about 1ppm for a 1000V => 10V transfer.

In a first step, you would measure 10V output from the 5440B on the 3458A, and determine a calibration factor between both, to about 0.2ppm uncertainty.
Measuring the 1V output voltage of your 1000:1 divider in the 10V range of the 3458A can be accomplished with < 0.5 ppm uncertainty, as described in the hp journal 4/1989, and also following its transfer / linearity specification (0.05 ppm of reading + 0.05ppm of range).

It is important to cancel out thermocouple voltages, which is possible in this case, as it's always a ratio measurement, not an absolute voltage measurement.

So apply 0V from the 5440B, zero the offset of the 1000:1 dividers output at the 3458A.
Then apply 1000V, let settle for at least 1min, and read the ~1V output voltage.

The best estimate for this 1000:1 transfer then would be 1ppm + 0.2ppm + 0.55ppm ~ 2ppm.

If you are a bit more conservative, you'd use the known specifications of the 5440B , that is 2.5ppm for 1000V, yielding about 3.5ppm in total.

Anyhow, I would guess, that your DIY 1000:1 divider is much less stable over time and temperature / applied power than that.

This can also be tested by this setup.

Frank

[e61_phil]

This whole calibration deals with ratio uncertainty and offsets, not absolute voltage uncertainty.
Therefore, many of the arguments about the real uncertainty of MisterDiodes do not apply here.

The ratio uncertainty of the 5440B (by its internal autocal process) is not specified at all.
It's only a guess from the addendum, where the autocal feature is described in prosa.
The tricky part of this description is, that for an external re-calibration it is sufficient to calibrate the 5440Bs 10V range only, and then do an internal autocal, omitting the ratio measurements by means of an 752A.

That is a point I've never understood. Why should I do a 100V and a 1000V adjustment, when the 5440B is throwing that away after AutoCal?

So, the guess would be, that autocal brings the instruments ranges relative to each other, back to the 30day specification, within 0.1ppm, as changes in ratios can be detected at that level.

If you substract the basic 10V uncertainty of 1.5ppm, which mainly describes the 30 day reference drift, from both range specifications, you may estimate the
100:1 ratio uncertainty between 10V and 1000V ranges (F.S.) to be around 1ppm. From the 5442A specification, this would even be 0.5ppm.

I also confirmed that with my own Hamon divider, which I calculated to be precise to about 1ppm for a 1000V => 10V transfer.

Sounds very good

In a first step, you would measure 10V output from the 5440B on the 3458A, and determine a calibration factor between both, to about 0.2ppm uncertainty.

That's the way I did it. I measured the 10V of the 5440B and the 1V output of the divider. How do you assume these 0.2ppm?

Measuring the 1V output voltage of your 1000:1 divider in the 10V range of the 3458A can be accomplished with < 0.5 ppm uncertainty, as described in the hp journal 4/1989, and also following its transfer / linearity specification (0.05 ppm of reading + 0.05ppm of range).

Why do you think it is < 0.5 ppm? My assumption was 10V -> 0.05 ppm + 0.05ppm = 0.1ppm and 1V in 10V range -> 0.05ppm+0.5ppm (10V/1V x range) = 0.55ppm which adds to 0.65 pmm

It is important to cancel out thermocouple voltages, which is possible in this case, as it's always a ratio measurement, not an absolute voltage measurement.

So apply 0V from the 5440B, zero the offset of the 1000:1 dividers output at the 3458A.
Then apply 1000V, let settle for at least 1min, and read the ~1V output voltage.

Offsets are a good point which I hadn't in my bill. Thanks!

The best estimate for this 1000:1 transfer then would be 1ppm + 0.2ppm + 0.55ppm ~ 2ppm.

If you are a bit more conservative, you'd use the known specifications of the 5440B , that is 2.5ppm for 1000V, yielding about 3.5ppm in total.

Anyhow, I would guess, that your DIY 1000:1 divider is much less stable over time and temperature / applied power than that.

This can also be tested by this setup.

Frank

Testing the divider over time was my intention for this calculation. I want to calculate the uncertainty of the ratio measurement itself, before I interpret any number from measurements every month or so. Temperature will also be tracked of course. The TC of the divider was measured to 1.5ppm/K.

[Dr. Frank]

That is a point I've never understood. Why should I do a 100V and a 1000V adjustment, when the 5440B is throwing that away after AutoCal?

Correct. That contradiction in the addendum made me think, that FLUKE at that time was not yet ready to declare their 544x calibrator as being an AUTOCAL instrument, as they did first with their later 5700A.
Secondly, the 544x has no full AUTOCAL functionality, as this does not work on a completely unadjusted instrument.. the dynamic range of the internal zero ADC for the Hamon divider bridge is too small, I assume.
But once a 752A calibration is initially done, this autocal really should work as well.

That's the way I did it. I measured the 10V of the 5440B and the 1V output of the divider. How do you assume these 0.2ppm?

From the transfer specification, you would assume 0.1ppm. In practice, you have noise, which gives about 0.2ppm standard deviation for 10V transfers.

Why do you think it is < 0.5 ppm? My assumption was 10V -> 0.05 ppm + 0.05ppm = 0.1ppm and 1V in 10V range -> 0.05ppm+0.5ppm (10V/1V x range) = 0.55ppm which adds to 0.65 pmm

In the hpj 4/89, they assume a typical linearity of about 0.02 ppm, and from that they estimate a 10:1 transfer of 0.33pm (afair). That I meant with being < 0.5ppm.
From the more conservative specification, 10:1 transfer uncertainty would be 0.55ppm, which I simplified to be in the same ballpark. I think, your additional 0.1ppm is not to be taken into account. Where should that come from?

Testing the divider over time was my intention for this calculation. I want to calculate the uncertainty of the ratio measurement itself, before I interpret any number from measurements every month or so. Temperature will also be tracked of course. The TC of the divider was measured to 1.5ppm/K.

You may do two different timely tests:
The first one is what you already indicated: Measuring the ratio factor (at identical temperature) over long term. This will give you the timely drift of the dividers resistors.
The second one is a short time drift monitoring, on msec scale.
You may digitize the 1V output, starting directly from the application of 1000V (by means of a reed switch, for example) over the first minutes, and resolving 1msec. This will give you the initial drift by self heating, or the T.C. matching of your divider resistors.


Frank

[MisterDiodes]

Spoke to a colleague whose business it is to measure HV accurately (To verify medical X-Ray equipment), and posed Phil's and Dr Frank's thoughts on a DIY HV divider - and a realistic uncertainty of such a divider (not just the meters used).  Basically - any accurate HV divider would be tested only at a verified known working voltage stress - Measurements over time will tell you if the system is stable or not - tells you nothing about the exact division ratio from unknown HV.  Since it doesn't sound like you have access to verified known 10kV voltage, you'll have to add a relatively large uncertainty since your trying to test your divider at only 10% of the unknown experiment voltage.  That is normally considered very poor practice, but you have to improvise where you can if you don't have access to a HV divider.

His first thought is by the technique you're attempting you'd be doing pretty good to get to within 1 or 2% given the nature of HV > 3kV, not some 10's ppm - without directly testing your divider against a known 10kV source - but there are ways to reduce uncertainty.

In other words - you're probably not worried about the ppm uncertainty of your meters, because your bigger unknown is in the HV divider itself.  His suggestion is to tackle that bigger problem first.

The first suggestion is to make a Hamon divider 10:1, calibrate at say 1000V (to read 500V as close as possible in cal mode).  Now use that (with your 5440) as a second way to estimate your 10kV and see how your resistor construction is performing at high stress.  The idea here is that you are using the most accurate, least division ratio possible to get within range to your most accurate meter at 1000V.  That will give you another estimate on where your unknown 10kV is really at.  You can repeat that experiment with a 100:1 divider to gain more insight on how the resistor construction is working.

Second suggestion is to be careful before using something like a Caddock resistor string at HV without expecting some real accuracy degradation - if that's what you're using for HV leg.  Realize that if your using a thin flat ceramic package charged to 10kV you'll tend to have surface leakage and every corner - so until proven otherwise he would go with an uncertainty of 100~250ppm for every resistor floating above 3~5kV as an initial ballpark estimate, depending on layout, arrangement, spacing, etc.  There are other estimates for different resistor packages but avoid any construction with square sharp corners.  You won't know until you actually test against a verified known HV source.  Putting the whole assembly in a vacuum will lessen leakage, but makes self heating a real issue.  Your HV divider will be no more accurate than how clean it is.  Your divider is also going to be very humidity dependent so you might think about a conformal coating.  These are things you might have already thought about.

Third suggestion is if you have access to an accurate, well known 2.5kV / 5kV and 7.5kV source that will give you a way to test your divider performance as you work your way up to higher stress levels.  The upper end of -all- HV dividers will leak (some designs worse than others) and cause errors, and the only way to really verify uncertainty is to keep testing against accurate, known sources - especially to see humidity effects.

Fourth suggestion is if your serious about HV that's beyond the range of the meter, then yes you float the meter up towards the HV to a read section of your divider - but that brings it's own set of problems and requires building a special HV "raised ground ref" environment for the meter to live in.

[David Hess]

This is a self build divider out of many selected Caddock USF resistors (high voltage part) and vishay foil resistors on the low side. It is only used with high impedance inputs.

How did you compensate for the voltage coefficient of resistance of the Caddock resistors?

I have considered making a precision high voltage probe but decided that the only way to make it precision is to grade identical or the same type of high voltage resistors for voltage coefficient of resistance which is very feasible and then use them in a divider string so the voltage across each resistor is identical or the voltage coefficient of resistance of the high and low part of the divider cancel.

[MisterDiodes]

My friend suggests to not waste time trying to re-invent a less-than-perfect wheel here on a 10kV probe.  With all due love & respect, if you can afford meter calibration, then you probably want to consider this route - or something similar.  These guys offer a good product at a good value - .05% is their standard catalog accuracy divider probe (to 50kVDC)  which is pretty good for a kV divider - and that is probably much better and cheaper than you're going to DIY with cobbled-together discrete resistors that aren't designed for operating at kV float levels:

https://www.cpshv.com/products/250.htm

They will also characterize your probe at say 10kV if you need for better accuracy at that cardinal point - but .05% for 50 kVDC divider probe is nothing to sneeze at for ~$500 USD, and their quality-check instruments are NIST-traceable cal'd to kV levels - which is not common.  A 70kVDC model is coming soon. 

These probes are available as true dividers setup for 10GOhm input bench meter or for 10M standard DMM meter - your order which one fits your needs.  Talk to them if you need more accuracy at 10kV.

They also make a nice 10kV benchtop supply that's handy if you're working on kV level experiments.

I would start there, and then see if it even makes any sense to spend the time and money on a DIY divider that would be hard to test if you don't have an accurate 10kV or higher source to test against.

Accurate kV measures are not a trivial task, that's for sure. 

[MisterDiodes]

Another option - this won't fit everyone's need but just FYI:

For less than $4000USD you can get an NIST-cal'd meter, with cal data (Cal data runs around $300 more over $3500 base price, I think), accurate to .03% basic DC accuracy.  Reads to 10kV directly, up to 140kV with additional probes.  Plus you get AC capability - which is -very- hard to DIY accurately:

https://vitrek.com/4700-precision-high-voltage-meter/

That's an excellent value if your looking for low measurement uncertainty at 10kVDC - especially if labor time spent building and verifying a self-built divider counts against business profits.

The fact that even high-end HV probes & meters stop at around .05% or .03% basic accuracy should give you a clue that it's going to be harder to build better accuracy. 

I'm not saying it can't be done, but it's going to be very hard to test and verify any DIY HV divider probe assembly to low uncertainties like we can do at lower voltages - just because of that HV is so hard to contain without errors.

The point is - if the divider probe isn't better than .05%, it doesn't matter too much if the meter used for measure is low ppm or 10's of ppm accurate... AutoCal / Ratio or whatever methods might not make much difference in the end.  Just depends on the application.

Have fun!

[e61_phil]

Why do you think it is < 0.5 ppm? My assumption was 10V -> 0.05 ppm + 0.05ppm = 0.1ppm and 1V in 10V range -> 0.05ppm+0.5ppm (10V/1V x range) = 0.55ppm which adds to 0.65 pmm

In the hpj 4/89, they assume a typical linearity of about 0.02 ppm, and from that they estimate a 10:1 transfer of 0.33pm (afair). That I meant with being < 0.5ppm.
From the more conservative specification, 10:1 transfer uncertainty would be 0.55ppm, which I simplified to be in the same ballpark. I think, your additional 0.1ppm is not to be taken into account. Where should that come from?

I applied the transfer specification twice. One time for the 10V measurement (0.05ppm + 0.05ppm = 0.1ppm) and a second time for the 1V measurement (0.05ppm +0.5ppm=0.55ppm). This is the way as it is described in the application note from Keysight.

Testing the divider over time was my intention for this calculation. I want to calculate the uncertainty of the ratio measurement itself, before I interpret any number from measurements every month or so. Temperature will also be tracked of course. The TC of the divider was measured to 1.5ppm/K.

You may do two different timely tests:
The first one is what you already indicated: Measuring the ratio factor (at identical temperature) over long term. This will give you the timely drift of the dividers resistors.
The second one is a short time drift monitoring, on msec scale.
You may digitize the 1V output, starting directly from the application of 1000V (by means of a reed switch, for example) over the first minutes, and resolving 1msec. This will give you the initial drift by self heating, or the T.C. matching of your divider resistors.

I used this self-heating test you mentioned to select the resistors.

But for clarification: I do NOT asume this ratio measurement as the complete uncertainty for high voltage measurements. My point here is simply to get a reliable uncertainty for exactly the measurement as described here. I will use this to learn something about the long term drift of the divider. And the measurements I've done in the past are all within lets say 10ppm. Therefore, it is neccessary to think about values in this range of precision.

Spoke to a colleague whose business it is to measure HV accurately (To verify medical X-Ray equipment), and posed Phil's and Dr Frank's thoughts on a DIY HV divider - and a realistic uncertainty of such a divider (not just the meters used).  Basically - any accurate HV divider would be tested only at a verified known working voltage stress - Measurements over time will tell you if the system is stable or not - tells you nothing about the exact division ratio from unknown HV.  Since it doesn't sound like you have access to verified known 10kV voltage, you'll have to add a relatively large uncertainty since your trying to test your divider at only 10% of the unknown experiment voltage.  That is normally considered very poor practice, but you have to improvise where you can if you don't have access to a HV divider.

His first thought is by the technique you're attempting you'd be doing pretty good to get to within 1 or 2% given the nature of HV > 3kV, not some 10's ppm - without directly testing your divider against a known 10kV source - but there are ways to reduce uncertainty.

I absolutely agree with you and your colleague, in the point that you cannot assumme accuracy in the range of 10s of ppm with such a measurement for voltages above 1kV. But 1 or 2% is far from realistic. I've build severeal HV-dividers and also a drift measurement system to measure drift in the sub-ppm range up to 10kV. This unit uses also selected Caddock Resistors (<1ppm/K) in a 16:1 Hammon configuration for calibration. This unit is temperature stabilized to a stability of about 10mK. I've measured severeal voltages between 0 and 10kV with the DIY divider which is discussed here and this unit. All measurements are within 5ppm of each other. I don't want to say that will mean it is accurate within 5ppm. But, for me, it is a strong hint that both units will measure much more accurate than 1 or 2%. Also commercial units, like the Brandenburg 149 (0,1%), agree to all measurements far better than the 0,1% spec of the Brandenburg.

And again: This uncertainty calculation as discussed here will be used to monitor the long term drift at 1kV.


And I don't apply 10kV to a single Caddock Resistor. In this case are 20 resistors in series. This results in a maximum voltage of 500V per resistor. Which reduces selfheating and voltage coefficient problems. And there is plenty of creapage distance on this ceramic resistor loaded only with 500V max. Caddock specifies 0.02ppm/V VC max. which is 10ppm at 10kV. And my experience is, that Caddock specifications are really conservative.

This is a self build divider out of many selected Caddock USF resistors (high voltage part) and vishay foil resistors on the low side. It is only used with high impedance inputs.

How did you compensate for the voltage coefficient of resistance of the Caddock resistors?

My approach is to use many resistors to reduce the voltage on a single resistor.

I have considered making a precision high voltage probe but decided that the only way to make it precision is to grade identical or the same type of high voltage resistors for voltage coefficient of resistance which is very feasible and then use them in a divider string so the voltage across each resistor is identical or the voltage coefficient of resistance of the high and low part of the divider cancel.

Sorry, but I don't get your point. (Sorry, my english is really bad )

My friend suggests to not waste time trying to re-invent a less-than-perfect wheel here on a 10kV probe.  With all due love & respect, if you can afford meter calibration, then you probably want to consider this route - or something similar.  These guys offer a good product at a good value - .05% is their standard catalog accuracy divider probe (to 50kVDC)  which is pretty good for a kV divider - and that is probably much better and cheaper than you're going to DIY with cobbled-together discrete resistors that aren't designed for operating at kV float levels:

https://www.cpshv.com/products/250.htm

Thanks, I know CPS. And I already asked them for 100ppm accuracy (which is the goal at the end). But no reply until now.

The other thing is, I would like to learn something doing this

Another option - this won't fit everyone's need but just FYI:

For less than $4000USD you can get an NIST-cal'd meter, with cal data (Cal data runs around $300 more over $3500 base price, I think), accurate to .03% basic DC accuracy.  Reads to 10kV directly, up to 140kV with additional probes.  Plus you get AC capability - which is -very- hard to DIY accurately:

https://vitrek.com/4700-precision-high-voltage-meter/

That's an excellent value if your looking for low measurement uncertainty at 10kVDC - especially if labor time spent building and verifying a self-built divider counts against business profits.

Thanks, I already had a flyer from that unit on my desk. The world of precision HV seems to be small . Two years ago I was searching something to measure the stability of voltages from -10kV to 10kV within at least 1ppm over 24h. CPS and Vitrek came across, but their specs didn't fit (accuracy wasn't important).

[ap]

See also: 'High Voltage Divider Calibration with the Reference Step Method', Speaker/Author: Harold Parks, National Research Council Canada

[Dr. Frank]

Why do you think it is < 0.5 ppm? My assumption was 10V -> 0.05 ppm + 0.05ppm = 0.1ppm and 1V in 10V range -> 0.05ppm+0.5ppm (10V/1V x range) = 0.55ppm which adds to 0.65 pmm

In the hpj 4/89, they assume a typical linearity of about 0.02 ppm, and from that they estimate a 10:1 transfer of 0.33pm (afair). That I meant with being < 0.5ppm.
From the more conservative specification, 10:1 transfer uncertainty would be 0.55ppm, which I simplified to be in the same ballpark. I think, your additional 0.1ppm is not to be taken into account. Where should that come from?

I applied the transfer specification twice. One time for the 10V measurement (0.05ppm + 0.05ppm = 0.1ppm) and a second time for the 1V measurement (0.05ppm +0.5ppm=0.55ppm). This is the way as it is described in the application note from Keysight.

OK, but you're not using the 10V and the 1V range to make the transfer, you're just using the 10V range to measure 10V for the 5440B output, and 1V from the divider output.
Therefore, the transfer specification only applies, instead of the two DCV specifications for measuring absolute voltages.

Frank.

[e61_phil]

I applied the transfer specification twice. One time for the 10V measurement (0.05ppm + 0.05ppm = 0.1ppm) and a second time for the 1V measurement (0.05ppm +0.5ppm=0.55ppm). This is the way as it is described in the application note from Keysight.

OK, but you're not using the 10V and the 1V range to make the transfer, you're just using the 10V range to measure 10V for the 5440B output, and 1V from the divider output.
Therefore, the transfer specification only applies, instead of the two DCV specifications for measuring absolute voltages.

Frank.

No, I used the transfer specification of the 10V range only.

The specification says: 0.05ppm of value + 0.05ppm of range. At 10V this is simply 0.1ppm, but at 1V the 0.05ppm of 10V (which is the range) will result in 0.5ppm of 1V. Therefore, you get 0.55ppm for the 1V measurement in the 10V range. This adds up to 0.65ppm.

[Kleinstein]

The construction of the divider can be tricky: with many resistors in series there can be still high fields toward the outside that could cause problems like corona discharge from corners. Though more important with AC dividers, even a DC HV divider might want to use a auxiliary dividers for shields to ensure a well defined field.

Such extra leakage is not well predictable and thus specifications need to be conservative. So typical readings can be much better, but it is really hard to get guarantied very accurate dividers. So a divider that could give you better tan 10 ppm error 80% of the time might still only be specified for 1 %.

[MisterDiodes]

...exactly.  Get above 3k or 5kv and you don't know what that first set of resistors in the string is doing; it all depends on humidity, cleanliness and construction details.  AC is a worse problem but you still get -plenty- of leakage at DC depending on the day.  Square corner Caddocks are a real challenge to keep kV potentials contained while floating at high kV, even Caddock will tell you that. To high kV, those resistor packages are just conductive plates with many sharp corner and long edge connections to the surrounding atmosphere - increasing the number of resistors increases the possible leakage nodes.

Not saying it can't be done of course, but no matter what it's a real challenge to keep high divider accuracy at high kV, especially if your kV source has high impedance.

Remember when Freon 114 wasn't dangerous and completely safe?  That's how we used to minimize divider resistor leakage when testing kV:

http://www.ebay.com/itm/Tektronix-P6015-High-Voltage-Probe-1000x/142470043982

See the can of freon?  You open the probe, squirt the liquid freon in to the fill line and close the probe up and let it pressurize - it will boil gently at room temp but under a little pressure it stays liquid.  Now your probe suddenly has much better dielectric qualities and maintains accuracy up to ~40kV.

[e61_phil]

Your are absolutely right, that high voltage is very tricky and AC HV is extremly tricky. BUT... I'm designing now since almost ten years high voltage circuits (power supplies, pulser, floating supplies...). And I earn my money with designing and building ultra-stable (single digit ppm over 48h) and precise high-voltage supplies, which are used for mass spectrometry. These supplies have to be stable and precise (tomorow, we will need the same voltage as yesterday within lets say 100ppm, but absolute accuracy doesn't matter). And the supplies uses feedback resistors, of course. If the feedback would act like you describe here, the ion signal would simply be gone away. In addition to that, the output impedance of such supplies must not be very low, to avoid producing lethal voltages.

And it is of course tricky and needs a lot of effort to bring a high voltage supply to that stability (coating, distances and so on. You said that already). The feedback resistors are problematic and need a lot of effort, but there are other problems which are greater.
My experience with high voltage is by far not as horrible as you describe it here.

Absolute accuracy wasn't a topic until now, because everything is tuned with ions. I'm also relativy new to metrology topics. Therefore, I asked the questions from the start. And I got great answers.  Thanks, again.


By the way: Today I tested the calibration of a HV divider I've calibrated in January. This divider has drifted around -22ppm in 8 months. Just to tell you, why I'm interested in uncertainty in such low levels. I've never saw severeral percent. But all dividers are carefully build. No sharp edges (metall balls at the junctions) and enough tracking distance. Everything is in a shielded box and so on.


PS: And I know the Tek 6105, of course . I own one of the old ones and at work we have also the newer ones (6015A) which don't need any fill.

[PartialDischarge]

PS: And I know the Tek 6105, of course . I own one of the old ones and at work we have also the newer ones (6015A) which don't need any fill.

Out of curiosity here's an X ray of the newest model P6015A, I believe it is filled with silicone gel, just like IGBTs,



Notice the darker area around the base, where the BNC connector goes. This is the low voltage side, where field gradients are basically quite low, and it seems like a void was created there to allow for the thermal expansion of the gel. Here's a CT scan I got. On the center there is the 100M resistor, and to the sides of it the distributed compensating capacitors made with metallic bars.

https://youtu.be/DBEN2vIK5iQ


By the way, the inner workings of this probe are not are simple as they may seem by it looks. It looks simple but is somewhat tricky

[MisterDiodes]

Yes, both the 60015 and 6015a are "filled" probes, the older version filled with freon, the newer version with silicone compound.  It would be difficult to build an accurate HV divider without some way of increasing dielectric properties (and/or guarding systems) around the high leg resistor elements.  Standard atmosphere is so unstable around kV with unavoidable humidity & pollutant particle level changes, fine aerosols, etc. - it's always a real challenge to know actual HV divider relative ratio ppm accuracy / uncertainty minute by minute (at kV levels).

E61_Phil would have to be careful about measuring a low apparent drift between across 8 months (and I'm sure he is very careful ) with a single set of measures:  That leaves at least two possibilities: #1) The entire system was stable and had repeatable measures 8 months apart or, #2) The measured 22ppm could be the relative drift between the unknown voltage divider ratio and unknown actual relative drift of the unknown power source.

In other words (just as a thought experiment) consider the situation where over 8 months the power source drifted ~500ppm, and the divider output drifted ~478 ppm the same direction - you'd still wind up with an apparent measured drift difference of ~22ppm.   That type of situation is common especially at kV levels.  This is the hidden source of uncertainty error with "relative" measures with multiple unknowns in the measuring technique - and the difficulty of not having an accurate, known stable 10kV source to verify a divider probe's HV stability -at stress design limits-, minute by minute.

 A low "apparent drift" difference of two measures by itself doesn't tell you anything about the real relative drift of either the power source or divider, but it does compare the -difference- between the two unknowns...PLUS whatever drift the meter encountered (but that one is measurable in calibration).  The other problem is that the various mechanisms causing drift in the HV divider can also cause a similar drift in the unknown power source, especially with HV high source impedance.  We've seen that happen several times and you have to be ready for it.

Without additional tests using different techniques and methods to rule out the much more common situation #2, there is no doubt measuring 10kV to some 10's ppm is not at all easy to do.  I'm sure E61 employs several other methods and techniques to measure the relative change in the 10kV power source over time in order to get lower uncertainty and verify he's looking at a relative long-term power supply drift that is closer to possibility #1.

Again, no measure at 10kV is easy.