Multiple voltage reference data question

[Conrad Hoffman]

Let's say I have N voltage standards (5 or 6 in my case). I measure the difference between them over time and temperature. Is there a way to extract which standard(s) are drifting? I can get a gut feel from the graph, but there must be a proper mathematical way. It will probably seem obvious once I see it, but I'm drawing a blank.

[Tj138waterboy]

What program are you using to save the recorded values?

[Conrad Hoffman]

Long term project and I'm just recording values twice a day on paper. I'll put them in Excel when I have some more.

[Andreas]

Hello,

short answer: No.
Reason: if you have 5 standards you have always only 4 (independant) difference measurements.
The other eventually measured differences can be calculated out of the others.

To solve the equations you need 5 independant values so at least one calibrated absolute measurement.

I also had to learn that when doing the first calibrations of my LTZ#1 and LTZ#2.
I knew from my difference measurements that the difference had drifted by around 12 uV.
But to my surprise LTZ#1 had drifted by -15uV and the other by -26uV between the first two calibrations.
So both much more than the 12 uV.

What you can do:
Detect which of the Standards is drifting more than the average value.
So sorting out the "stinkers".

with best regards

Andreas

[CalMachine]

Let's say I have N voltage standards (5 or 6 in my case). I measure the difference between them over time and temperature. Is there a way to extract which standard(s) are drifting? I can get a gut feel from the graph, but there must be a proper mathematical way. It will probably seem obvious once I see it, but I'm drawing a blank.

Conrad,  this might help 

http://www.dataproof.com/media/2f556cab2ad70fdffff821affffe905.PDF

It's the mathematics behind a lot of volt reference scanning software

[SilverSolder]

Let's say I have N voltage standards (5 or 6 in my case). I measure the difference between them over time and temperature. Is there a way to extract which standard(s) are drifting? I can get a gut feel from the graph, but there must be a proper mathematical way. It will probably seem obvious once I see it, but I'm drawing a blank.

Conrad,  this might help 

http://www.dataproof.com/media/2f556cab2ad70fdffff821affffe905.PDF

It's the mathematics behind a lot of volt reference scanning software

Link not working.   Any chance of attaching the PDF to a message here instead?

[BravoV]

Let's say I have N voltage standards (5 or 6 in my case). I measure the difference between them over time and temperature. Is there a way to extract which standard(s) are drifting? I can get a gut feel from the graph, but there must be a proper mathematical way. It will probably seem obvious once I see it, but I'm drawing a blank.

Conrad,  this might help 

http://www.dataproof.com/media/2f556cab2ad70fdffff821affffe905.PDF

It's the mathematics behind a lot of volt reference scanning software

Link not working.   Any chance of attaching the PDF to a message here instead?

Working just fine here, attached below.

[dietert1]

Still observing multiple references is better than observing two when it comes to predictions/extrapolations for the future.
The problem arises when you think the other references will make predictions about the future of one of the group. This isn't possible. By observing multiple references you will likely detect technical problems with one of them (for example the one you built/bought last). But you won't be able to predict such problems.

From some Fluke document i remember: They reduce error bars on the predictions for their references after they have been regularly calibrated (compared to others) for several years. If a reference exhibited a certain behavior for several years (e.g. moderate continuous drift) it will likely show a similar behavior in the future. That seems to be a valid idea, as long as certain requirements on handling the references are observed.

Regards, Dieter

[Conrad Hoffman]

Thanks! That NBS paper is good. I've got a lot of NBS/NIST info on standard cells and measurement, but I don't think I've seen that one. I think what I was expecting was to identify "stinkers", as there's no way to establish whether the whole group has drifted together. At one time I had an ovenized standard cell bank but got rid of it because I didn't want mercury in the house. In hindsight, I should have kept it as the risk was minimal and it was extremely stable over a long time period. OTOH, it was a pain maintaining and replacing the batteries, which are absolutely necessary because of the recovery time should the temperature change.

[Andreas]

Hello,

here a Diagram of Drift of difference voltages of 4 References.
LM399#1 + LM399#2
LTZ1000#1 + LTZ1000#2

X-Axis: day (so roughly 3 kHrs)
Y-Axis: drift of the difference voltage in micro-Volts.

So the only conclusions I can make:
All differences with LM399#1 "drift as hell" (so this one is a stinker).
The difference between LTZ#1 and LTZ#2 is "quieter" than the other differences.

With best regards

Andreas

[notfaded1]

@Andreas it also makes it look like generally the LTZ are pretty superior! 

Bill

[dietert1]

When i look at the (violett) difference between the two LTZs, between day 30 and 60 there is an almost perfect run, where both LTZs agree very well. After 60 days one would estimate that they agree within 1 ppm, so about 6 ppm for a year (simple minded extrapolation).
I can also see something like 3 ppm around day 110. It doesn't appear like a moderate continuous drift. So an upper limit over a year would probably be around 10 ppm.

Maybe Andreas has more data. I mean nothing will protect you from an "event" that disturbs your voltage reference from one day to the other by a ppm or some ppms. It can be popcorn noise, a hard hit/drop, a power outage, overheating, component failure, a special cosmic ray event etc.

Regards, Dieter

[Andreas]

Hello dietert1,

I would not interpret too much into the measurements in uV resolution with a 24 bit ADC with 2:1 voltage divider and 10V range in total.
There where several uncertainities when I started those measurements.
- noise of the references (LM399#2 is rather noisy compared to other LM399´s that I have).
- noise of the 24 Bit ADC with day to day (peak-peak) variations up to 4 uV in 10V range. (later I could reduce this noise to 2 uVpp/10V)
- zero offset drift due to bad connector contacts up to 4uV/5V so 8uV/10V (which was later solved by deoxit; see day 2800 in diagram)
- and last but not least: the 2 LTZs difference of 40mV looks better than the ~300mv difference between a LM399 reference and a LTZ reference because there is less (absolute not relative) gain error by the ADC voltage reference (at that time it was mainly due to humidity changes of the LT1027)

But it is always good to have a starting point to gain experience.

With best regards

Andreas

[Conrad Hoffman]

Unlike a lot of circuits, every reference I've looked at, even with identical construction, has visibly different noise.

Way back when I did the Mini-Metrology series, I compared the voltage reference with my standard cell bank. I did this almost every day for two years. The data was recorded in an excellent old spreadsheet called Javelin. It was designed for financial work and could handle time series better than most anything. Alas, it was DOS and couldn't run on anything modern. Software and data are long gone. https://en.wikipedia.org/wiki/Javelin_Software

I'm thinking the time is near to build my own LTZ board.

[Grandchuck]

I'm thinking the time is near to build my own LTZ board.

That is what I am doing.  What are your plans regarding suitable resistors?

[hwj-d]

booked

[Conrad Hoffman]

No solid plan yet! I'd probably have to order resistors. I have a good collection of wire-wounds, but they're nowhere near as good as the Vishays. I could also hand select metal films for value and tempco. My previous experience with that was good, but I don't know about fractional ppm stability.

I keep thinking about the math for stinkiness. It seems like if one takes a given time period and does an RMS of the values for each unit, vs the RMS for the group, that should give the stinky factor. Probably need to subtract an offset for each one, since the average DC value isn't what I'm after.

[Kleinstein]

To check a set of references for bad ones, one can look at the difference between the average and the individual references.
The critical part is usually the lower frequency noise (e.g. < 10 Hz). So one can use relatively slow readings and do RMS calculation without the DC background in software.

The good thing about the LTZ1000 circuit is that the OPs and resistors are not that critical. There errors and noise are attenuated by about a factor of 100. The critical part is usually long term stability not so much the TC or noise.