EEVBLOG #210 Krohn-Hite Voltage Standard

[saturation]

Yes, my bad, error corrected.  white noise, statistical mean = 0, variance finite.

Yes, during a 24 hour period, its unlikely a marked shift will occur to the average of environmental variables, for example from the weather graph dates Nov 27-29 and Dec 3 have less variable temps, humidity, pressure, but different averages for each group, not including the diurnal shifts.  For a weeks worth of data, measures includes the Nov 29-Dec 2 rising barometric pressure as well as falling average temps and humidity.   I've found it only affect uV levels and smaller.

Yes, temperature is undoubted a biggest factor, but its one we can do something about, so climate control in a room will minimize larger variance, but trying to stabilize the last few LSD is a challenge.  Assuming we've put controls on other things like triboelectric, electrochemical, guarding , etc., 

Yes, I've considered it. But as you showed, if the shorted output does represent 3456a internal drift + noise, it only contributes variation in single digit nV,  at the limits of the 3456a resolution.   I didn't have a reference stable below 1uV to give it a try, but with the MV106, I'll try.   

Its possible to estimate a correction similar to these folks [its free to download, hopefully for all]:

http://www.springerlink.com/content/725668231627q136/

I believe you're using variance in a non-technical sense, this is confusing in this context.

Of course both the DMM and the voltage reference will be affected by this. So taking samples at 24h intervals would give a lower variance? The drift appears to be dominated by temperature coefficient? If you could measure just the effect of the DMM, you could correlate the ambient temperature with the drift of the DMM, and attempt to correct for this. This is cheaper than to get a proper climate-controlled room .

Guess this is why 24h specs are for 23°C ± 1°C.

[alm]

Yes, temperature is undoubted a biggest factor, but its one we can do something about, so climate control in a room will minimize larger variance, but trying to stabilize the last few LSD is a challenge.  Assuming we've put controls on other things like triboelectric, electrochemical, guarding , etc., 

As long as the noise is uncorrelated, averaging will help, so you shouldn't have to worry about some random fluctuations. This won't help against cyclic behavior like the daily temp/humidity/pressure changes, unless you set the sample period to be a multiple of 24h, just like how they set the integration time to be a multiple of a PLC to reject the mains noise.

Yes, I've considered it. But as you showed, if the shorted output does represent 3456a internal drift + noise, it only contributes variation in single digit nV,  at the limits of the 3456a resolution.   I didn't have a reference stable below 1uV to give it a try, but with the MV106, I'll try.   

Shorted output only represents noise and offset drift, not gain drift. Not sure how to characterize gain drift without taking the source into account, it's a lot easier to make good short than a good voltage source. Connecting multiple meters to the same source might help, although I'm not sure if their drift will be uncorrelated if they're the same model with the same LMx99 reference.

Its possible to estimate a correction similar to these folks [its free to download, hopefully for all]:
http://www.springerlink.com/content/725668231627q136/

Except for the constant pressure/temperature chamber, I'm guessing . You may be able to figure it out just from correlation if you collect enough data. At a first glance their model seems straightforward and quite usable in your case.

[EEVblog]

There are hardly any reviews of EDC products or discussions of them [ at least via google].  So, very likely this thread will pop up in the near future when EDC or Krohn Hite are searched for.  This device and its derivatives are worthy of more interest from anyone looking a well made, portable affordable voltage standards that can be priced right for small labs.

Already link #2 for "Krohn-Hite voltage standard", with the video at #1

Dave.

[saturation]

Good work Dave, your video and forum links pop up as the #1 item when you google "krohn hite voltage standard"  and among the first 10 using "krohn hite MV106 review."

I hope it helps them as it helps you.

The EDC equipment portfolio is not well documented on the web, if not for the KH website.  There are models of theirs that pop up now and then on eBay that have no documentation I can find.    There seems to be evolutionary models of the MV106J, such as the MV105 & MV100 that are identical to the 106 but less switchable voltage ranges.

They have several generations of AC calibrators that rival the then standard Fluke and HP models but are not huge boat anchors, even for its time, such as the CDC-32.   These models are not in the KH site.




FWIW, I'm now testing the MV106 stability at 1uV output and at just over 24h the statistical variance on the 3456a is amazing at 6 x 10^-15, or an SD of 77 nV.

There are hardly any reviews of EDC products or discussions of them [ at least via google].  So, very likely this thread will pop up in the near future when EDC or Krohn Hite are searched for.  This device and its derivatives are worthy of more interest from anyone looking a well made, portable affordable voltage standard that can be priced right for small labs.

Already link #2 for "Krohn-Hite voltage standard", with the video at #1

Dave.

[robrenz]

saturation, 
I find the metrology aspect of this thread is very interesting. (my metrology knowledge is mostly mechanical I have a long way to go in electronics) Your stability testing made me wonder how stable my 8846A is on DC volts. So I followed your lead and did a shorted input test for 12 hours overnight in statistics mode at 100NPLC after a 2 hour warm-up. No special temp control on the room, just thermostat with ceiling mounted forced air heater.  In stat mode the min/max resolution is 0.1nV . After 12 hours overnight with 26000 readings the SD was 180.979nV. I think that would be 0.181 ppm. Is that a good number for a meter of this level?
I have noticed that the 8846A works internally (on the lowest range only) at 1000X higher resolution than the displayed LSD (I do know the difference between resolution and accuracy). You can see the full readings in stat mode or trend plot. Max DCV resolution is 0.1uV but internally is 0.1nV, Max Ohms resolution is 10uOhm but internally is 10nOhm etc. That 1000X resolution is there  from 0.02NPLC to 100NPLC and with digital filter on or off.  It still only has 6.5 digits but on the lowest range but it seems to internally change amplification a decade at a time based on the reading to maximize resolution with a limit of 0.1nV.  Do all 6.5 digit meters do this?

Edit: corrected to 0.1nV  was 01.nV

[saturation]

Yes concur.  To reduce the effects of cyclic variations of daily weather and then climate with the seasons, I compare the Geller SVR reference output only when the weather conditions are identical to the calibration date.  Its essentially identical at 1 year, which is not what estimated predictions suggest it would be, once variations accounted for by weather are controlled.  Thus, I don't even need a make a coefficient to correct my reference voltage.

FWIW, the MV106 is far more immune to weather, as we've had a large change in conditions this week and the statistical variance it caused are minimal.  But can't say yet as have no longer term data as the season progresses.


Can you clarify what 'gain drift' refers too?  The internal amps of the DMM?

As long as the noise is uncorrelated, averaging will help, so you shouldn't have to worry about some random fluctuations. This won't help against cyclic behavior like the daily temp/humidity/pressure changes, unless you set the sample period to be a multiple of 24h, just like how they set the integration time to be a multiple of a PLC to reject the mains noise.

Shorted output only represents noise and offset drift, not gain drift. Not sure how to characterize gain drift without taking the source into account, it's a lot easier to make good short than a good voltage source. Connecting multiple meters to the same source might help, although I'm not sure if their drift will be uncorrelated if they're the same model with the same LMx99 reference.

Except for the constant pressure/temperature chamber, I'm guessing . At a first glance their model seems straightforward and quite usable in your case.

[saturation]

Hello robrenz,

On the bold items, yes, its superb.  I don't know if all 6.5 digits DMMs do this.  Its very possible your DMM has more precision internally than the readouts show at face value. 

A question is how much system noise or drift is created by the DMM's design to make that extra resolution useful?  Assuming your measurement connections were optimal and you got 0.18ppm, its fairly close to my 3456a.  However the 3456a resolution is 100nV [ it could be 10nV, see below]; so the Fluke's extra 0.1nV resolution could be limited by other issues and to use it requires much care and planning; its just a speculation.

IIRC, the 3456a may use internally to 7 digits at the 0.1V scale, so it may really have 10nV resolution instead of the publicized 100nV, but it can only be used via the GPIB port, analogous to what you have with the Fluke.  The Fluke 8846 competes against the defacto standard bench DMM, the Agilent 34401a, so its built to meet and beat its specs. 

The 34401a has many similarities to the 3456a, but its more modern and more portable compared to the rack mounted 3456a.  01a incorporates many measurement design concepts made in the 34xx series DVM: 3455, 3456, 3457, 3458.  The 3458a is a widely used 8.5 digit DMM often found in metrology labs.  The design concepts of these DMM were published in the HP journal when it was around so, its possible competitors have studied those techniques and use them in their designs.

saturation, 
I find the metrology aspect of this thread is very interesting. (my metrology knowledge is mostly mechanical I have a long way to go in electronics) Your stability testing made me wonder how stable my 8846A is on DC volts. So I followed your lead and did a shorted input test for 12 hours overnight in statistics mode at 100NPLC after a 2 hour warm-up. No special temp control on the room, just thermostat with ceiling mounted forced air heater.  In stat mode the min/max resolution is 0.1nV . After 12 hours overnight with 26000 readings the SD was 180.979nV. I think that would be 0.181 ppm. Is that a good number for a meter of this level?
I have noticed that the 8846A works internally (on the lowest range only) at 1000X higher resolution than the displayed LSD (I do know the difference between resolution and accuracy). You can see the full readings in stat mode or trend plot. Max DCV resolution is 0.1uV but internally is 0.1nV, Max Ohms resolution is 10uOhm but internally is 10nOhm etc. That 1000X resolution is there  from 0.02NPLC to 100NPLC and with digital filter on or off.  It still only has 6.5 digits but on the lowest range but it seems to internally change amplification a decade at a time based on the reading to maximize resolution with a limit of 01.nV.  Do all 6.5 digit meters do this?

[robrenz]

Thanks saturation,
I am restarting the test again to do at least 24 hours and I am recording the min/max temperature at the case during  the test.

robrenz

[alm]

Can you clarify what 'gain drift' refers too?  The internal amps of the DMM?

Amplifiers, dividers, internal voltage reference, ADC and anything in between. All will drift (very slightly). With shorted inputs, you're just measuring offset (which is usually compensated by auto-zero) and noise. Change in gain is irrelevant for this measurement, if your voltage reference would change from 10V to 9.9V, you wouldn't notice any difference with shorted inputs. Not saying I expect much for a well-designed DMM like the HP 3456A, but I can't imagine it'd be zero, just look at the 24h and tempco specs. Of course these are worst case. The gain error is specified by the % of value uncertainty; the offset error/noise is represented by the % of full scale error.

After 12 hours overnight with 26000 readings the SD was 180.979nV.

This is very close to the 170 nV saturation reported for n ~ 12000. The number of samples is sufficiently large that both should represent be a very good estimate of the population standard deviation, assuming it's purely determined by noise. Temperature shouldn't really matter for this measurement, although you might expect slightly more noise at higher temperatures. The difference between say 293K and 300K is unlikely to make much difference, I would expect. I believe the HP 3456A has better accuracy specs than the Fluke 8846A (it was a reference class meter in its day, similar to the 3458A now), not sure about noise.

I have noticed that the 8846A works internally (on the lowest range only) at 1000X higher resolution than the displayed LSD (I do know the difference between resolution and accuracy). You can see the full readings in stat mode or trend plot. Max DCV resolution is 0.1uV but internally is 0.1nV, Max Ohms resolution is 10uOhm but internally is 10nOhm etc. That 1000X resolution is there  from 0.02NPLC to 100NPLC and with digital filter on or off.  It still only has 6.5 digits but on the lowest range but it seems to internally change amplification a decade at a time based on the reading to maximize resolution with a limit of 01.nV.  Do all 6.5 digit meters do this?

Some do, especially Keithley. A random result via GPIB from a 5.5 digit meter with shorted inputs on its 300mV range (expect resolution 1 uV)
+000.0042E-3 (resolution .1 uV)

From a 6.5 digit meter on 100mV range:
4.75412547E-08 (.1 fV resolution!)

Especially the last one is completely ludicrous, I consider this just artifacts of the calculation and completely insignificant compared to the noise floor. The only advantage of the extra resolution is that it reduces quantization error, although it should be random for this (haven't studied it in that much detail). This is a real issue when averaging down to the last digit. Datron actually introduced noise in the measurement and then averaged it to get a better resolution. That's how they achieved 7.5 digits, I wouldn't be surprised if it was still used in the Fluke 8508, since they bought Datron to get their high accuarcy DMM technology). Most other brands return the same resolution via GPIB as on the display. Except for the 3457A which is specified as a 6.5 digit meter, but 7.5 digits of resolution via GPIB. This does not translate in improved accuracy, however, since the 3456A is superior in that regard.

[robrenz]

I have noticed that the 8846A works internally (on the lowest range only) at 1000X higher resolution than the displayed LSD (I do know the difference between resolution and accuracy). You can see the full readings in stat mode or trend plot. Max DCV resolution is 0.1uV but internally is 0.1nV, Max Ohms resolution is 10uOhm but internally is 10nOhm etc. That 1000X resolution is there  from 0.02NPLC to 100NPLC and with digital filter on or off.  It still only has 6.5 digits but on the lowest range but it seems to internally change amplification a decade at a time based on the reading to maximize resolution with a limit of 01.nV.  Do all 6.5 digit meters do this?

Some do, especially Keithley. A random result via GPIB from a 5.5 digit meter with shorted inputs on its 300mV range (expect resolution 1 uV)
+000.0042E-3 (resolution .1 uV)

From a 6.5 digit meter on 100mV range:
4.75412547E-08 (.1 fV resolution!)

Especially the last one is completely ludicrous, I consider this just artifacts of the calculation and completely insignificant compared to the noise floor. The only advantage of the extra resolution is that it reduces quantization error, although it should be random for this (haven't studied it in that much detail). This is a real issue when averaging down to the last digit. Datron actually introduced noise in the measurement and then averaged it to get a better resolution. That's how they achieved 7.5 digits, I wouldn't be surprised if it was still used in the Fluke 8508, since they bought Datron to get their high accuarcy DMM technology). Most other brands return the same resolution via GPIB as on the display. Except for the 3457A which is specified as a 6.5 digit meter, but 7.5 digits of resolution via GPIB. This does not translate in improved accuracy, however, since the 3456A is superior in that regard.

It seems from what I have tested so far that the extra 3 digits are real not calculated. It is the same at .02NPLC and no digital filtering so I would think in that case it is actually a single reading. This shows the 0.1nV resolution showing on the stat display The max and min are not calculated values and the last two trailing zero's are always there.  The avg. and Sd are calculated and show more digits.  Keep in mind as soon as the max or min value needs another most significant digit to display the value, the resolution will decrease one digit.  So this meter only does this as values approach zero. As soon as 6.5 digits are required to display the value, the stat display will have the same resolution as the main display.

[robrenz]

I talked to Fluke tech support and the 8846A does in fact do autoscaling to maximize the available internal resolution past the stated max main display resolution.  It does this anywhere in the range not just around zero. So if you hooked it up to a precision 10V source and the source only deviated a few uV it would autoscale to the max resolution on the stat or trend plot displays. It will always use the max internal resolution possible. I have not tried this yet at a voltage other than zero but as soon as I am done the 24hr stability test is done I will.

[saturation]

As an aside here's interesting cal lab cumulative report on the stability of multiple 34401a compared to  Keithley 2000s.

There is a public discussion on a metcal technician forum about the Fluke 8845a, most of their concerns were about input jack quality.

http://www.pmelforum.com/index.php?topic=1429.15

[alm]

I talked to Fluke tech support and the 8846A does in fact do autoscaling to maximize the available internal resolution past the stated max main display resolution.

Any spec for this internal resolution?

  It does this anywhere in the range not just around zero. So if you hooked it up to a precision 10V source and the source only deviated a few uV it would autoscale to the max resolution on the stat or trend plot displays. It will always use the max internal resolution possible. I have not tried this yet at a voltage other than zero but as soon as I am done the 24hr stability test is done I will.

As I stated, the value is somewhat limited if the noise floor is ~0.2 ppm, apart from reducing quantization error.

As an aside here's interesting cal lab cumulative report on the stability of multiple 34401a compared to  Keithley 2000s.

Shows why the 34401A has been the industry standard forever. Funny how the far majority of the out of tolerance points for the Keithley are in the resistance ranges, increasing with the higher ranges, but dropping again in the 10/100 Mohm ranges, probably because of the looser tolerances.

[robrenz]

I talked to Fluke tech support and the 8846A does in fact do autoscaling to maximize the available internal resolution past the stated max main display resolution.

Any spec for this internal resolution?

Max resolution is 0.1nV but it never exceeds 6.5 digits of information. You can see an example in the picture of the screen in a post below.

[robrenz]

8846A results from a 24 hour test of shorted input DcV   Much better than the first 12 hour test
Max. reading -2.9535uV  Min. reading -1.8006uV  (zero does not function in stat mode these include the meter offset)
total 24 hour span 1.1529uV
@ 1hour 2064 readings SD=60.96nV
@ 6 hours 12387 readings SD=83.47nV
@ 12 hours 24774 readings SD=86.1nV
@ 24 hours 49000 readings SD=133.32nV

Air temperature at case during test: Max. 30.89 degC Min. 22.67degC  8.2 degC temp variation.  (I forgot about my thermostat setting back the temp overnight)
Will I see a big difference if I run again with the temp more stable? (no temp setback overnight)

[saturation]

Thanks robrenz for making this available.  As expect, its definitely doing as well, if not better, than the 3456a. 

In the end, this exercise is just to get insight into how the the whole system drifts by monitoring the offset voltages, and using those results to guestimate how the DMM drift will contribute errors when measuring a high quality voltage reference.

As for the effect of setting your thermostat, its unknown until you try, it should be easy to do if you have time and and it could settle any questions you have of its stability. 

At 6.5 digits, we are edging to the limits of DMM design.

8846A results from a 24 hour test of shorted input DcV   Much better than the first 12 hour test
Max. reading -2.9535uV  Min. reading -1.8006uV  (zero does not function in stat mode these include the meter offset)
total 24 hour span 1.1529uV
@ 1hour 2064 readings SD=60.96nV
@ 6 hours 12387 readings SD=83.47nV
@ 12 hours 24774 readings SD=86.1nV
@ 24 hours 49000 readings SD=133.32nV

Air temperature at case during test: Max. 30.89 degC Min. 22.67degC  8.2 degC temp variation.  (I forgot about my thermostat setting back the temp overnight)
Will I see a big difference if I run again with the temp more stable? (no temp setback overnight)

[saturation]

Amp-hour posted this notice, Keithley's latest text on the how-and-why of low level measurements is available free:

http://www.keithley.com/data?asset=9538

If the link fails try the main search:

http://www.keithley.com/at/027.html

Search for: Low Level Measurements Handbook

[robrenz]

They will send you a paperback version for free also. I got one 6 months ago.
http://www.keithley.com/knowledgecenter

[saturation]

A manual for the MV106 and other venerable older equipment are available here:

http://www.ko4bb.com/manuals/index.php

[Galaxyrise]

I'm playing with an EDC 2902 now, built in 12/79, which seems to be based on a very similar design.  Documentation has been sparse, and the video was great at providing some insights for me.  Thanks, Dave

I think the mystery device is a precision 3 ohm resistor. 4 terminal so two terminals can be used for the series current through the resistor, and two outputs to provide the precise 3 ohm output without any lead resistance error.

My mystery device is marked A2405-1, unlike Dave's which is A3550.  A3550 does appear on Dave's schematic by the 3R resistor, and between OUT and COM, I do measure 2.984.  But between IN and OUT, I measure 296.8.  between COM and IN, I measure 299.8.  So I think the A3550 is both the 3R and the 297R.

Thanks to the video, I know what the label on the right is referring to, yay   (It's been cut off, but it reads 1893, 6.372, @7.5) .  However, in my unit, the zener looks a LOT different than Dave's; I mistook it for a resistor at first!  (The little black glassy component under the yellow wire, banded with grey white yellow orange.) Stability of mine is pretty poor compared to modern voltage references, and I may try replacing it with something else.

I can also offer some insight as to why Dave's chopper amp was on a daughter board.  Check out the amp on my board!  It's got the same pin arrangement on the bottom as Dave's has.  Another possible project for me

All the diodes in this unit (and only the diodes) have a U-bend added to the lead on the cathode side. Some kind of strain relief, I presume?