Advantest R6581 8.5 digit DMM mini teardown/repair

[quarks]

I do not know the exact year of my R6581, but most IC datecodes are from 1995 and 1996, therefore it was most likely build in 1996

[quarks]

there is a stamp "9640" on some of my boards
If your boards are newer, there must have been an change and the C215 Problem was not solved, but instead created/introduced with this change.

[branadic]

Hopefully someone can help interpreting the following results...
I hade Datron4000A, K3458A and R6581D in our big climate chamber. I first set temperature to 23°C @ 45%rH and waited over 24h for everything to stabilize. I then performed internal cal on R6581D and ACAL on K3458A, then started my frist INL measurement.
Afterward I've set temperature to 20°C @ 45%rH and repeated the INL measurement without new internal CAL / ACAL. Finally I repeated the measurement at 26°C @ 45%rH. Attached are the results I got.
First row is K3458A deviation (blue) and R6581D deviation (cyan) vs. Datron 4000A. (K3458A - Datron4000A & R6581D - Datron4000A).
Second row is R6581D deviation vs. K3458A.
Third row is temperature vs. Datron4000A, simply to show that temperature was stable during measurement.

What is clearly visible is, that I was able to perform INL measurement at a given temperature without influence of temperature variation during that time (-11V up to +11V and from +11V down to -11V, so two full "ramps").
R6581D matches Datron4000A best at 20°C with a slight offset (flat INL curve), while K3458A says that both show a gain error. The shape of the error curve for R6581D looks to be constant, though influenced by a temperature dependent gain error.

What else can I extract? Unfortunately I didn't yet managed to perform temperature of K3458A and R6581D against a constant reference voltage, but maybe that isn't necessary as the information is already captured with the measurements I already performed?

-branadic-

[Kleinstein]

The 3458 vs Datron4000 seem to be off a little in the absolute value (e.g. calibrated from different sources). Some 3-4 ppm is not very much and there seem to be a slightly different TC too. One the good side the linearity seem to be quite good.


It looks like the R6581 has some gain TC relative to the 2 other instruments - still in the 1ppm/K range.

The INL error looks a little like some of the curves show a clear pattern, that does not change much with temperature.
There are 3 sections: -10 to -4 V, -4 to 0 V and the positive side.
The jump (some 0.6 ppm) near zero is nothing really new - chances are high it is related to DA as there is also a jump in the average integrator output voltage near zero.  This part is probably the most troublesome part, as the internal AZ measurement uses a conversion near zero. If this just happens to be in the steep step part this can increase the errors quite a bit.

edit:
For the DA the averaged integrator voltage is a good indication of what to expect. User Mickle has measured and shown this here:
https://www.eevblog.com/forum/metrology/advantest-r6581-8-5-digit-dmm-mini-teardownrepair/msg1535345/#msg1535345
So I would interpret the curvature for positive voltages and the jump as caused by DA. The break in the curve near -4 V  may be due to another effect.

For the integration cap, there may be the possibility to get lower DA from good NP0 capacitors. For the TDK brand  I found DA at about 1/5 of the PP ones I tested. It is still not sure how good or bad the installed original cap is. The actual DA can vary between brands / types (e.g. purity of the PP material). As they claim a much better INL, chances are they originally had really good  caps. Possibly the caps just got worse with age or the later caps were just not as good and nobody noticed this. Due to the slow modulation during run-up the ADC design is definitely very sensitive to DA.  When looking for good caps, keep in mind the usually DA value give is for some 10-900 seconds. The relevant time range here is more like 2-200 ms. So the loss factor at 1 kHz may be the better approximation.  Anyway the numbers in the Data-sheets are usually limits, possibly just limits of there tests.

[dietert1]

What happens when you vary humidity e.g. 35 % RH and 55 % RH at 23 °C?

Dr. Frank recently emphasized how you have to watch ambient temperature of a 3458A DVM when it comes to the last ppm. Did anybody try to convert one of those cheap incubators to a temperature oven for a 19" DVM?

Regards, Dieter

[branadic]

As time in the chamber is limited I've started measurements of a 10V source (temperature compensated LTZ1000) at constant temperature, while temperature with DMMs inside the chamber is varied. Therefore I again performed ACAL/internal Cal at 23°C and 45%rH, measured some hours and set temperature already to 20°C. Will go back to 23°C, afterwards 26°C and again 23°C.
I guess there is not much time left as I need to leave chamber tomorrow.

My hope was, that Frank leaves a comment, too, but maybe he is still thinking about his answer?

-branadic-

[Dr. Frank]

Hello branadic,
That way, I can't extract no reasonable information.

these graphs show, that there is a gain (adjustment) error between the 3458A and the 4000A, as well between all three instruments.
So for an INL graph, you would need to normalize these gain (and offset) errors first, to get the INL over U.
What was your intention, in the end?
To determine the INLs of the 4000A and the R6581, assuming very low INL for the 3458A?

For the extraction of T.C.s that's also not so good, as all three T.C.s interfere, so your running T.C. measurement on the 3458A probably will let you solve for the other T.C.s
A wider temperature span like +/- 5 ..10 °C would have been better, I guess.

So a nice experiment, when you calculate the INLs correctly.

Frank

[branadic]

Thanks for the answers, that I will answer later, as I'm still busy in analyzing them.

@ Kleinstein
The capacitors used are 1.5nF for fast sampling and 20nF for slow sampling polypropylene by Soshin.
Compared to other capacitors of different brands they are specified for low dissipation and tolerance:

Film capacitors for general use NQP
Low loss (about 0.02%, 1kHz) capacitor.
A small tolerance of ± 1% has been realized.
Small and high precision capacitors with good space factor
Suitable for filter circuits.

By now I couldn't find a proper replacement for testing, as 20nF is somewhat unusual and has to be made by paralleling two 10nF capacitors. I couldn't find proper NP0/C0G capacitors from TDK, as suggested by you. Maybe you can name a capacitor series number?

-branadic-

[TiN]

Ran quick INL sweep on R6581T have here. Results for -10.9 to +10.9 run in 0.1V steps, NPLC30, soak time 10s between points:



3458A meeting promised 0.1ppm INL. GPIB3 is one of my reference units, GPIB11 is standard unmodified instrument. My broken R6581T shows INL error +0.28 to -0.20 ppm in this run, rather poor as for 8.5D DMM, but aligns with other owners.

As always, RAW-data available, no secrets here. Column source - programmed value, duta = 3458A GPIB3, dutb = 3458A GPIB11, dutc = R6581T.

Plot generated by python via matplotlib and numpy against polyfit for each instrument row. Config file for app.

[Kleinstein]

The capacitors I got the best DA with were TDK  C2012C0G1H222J085AA (mouser 810-c2012c0g...).
These are 2.2 nF, but I would expect slightly larger capacitance values to be similar. From the DA test I would expect a loss at 1 kHz on the order of 0.005%. The data-sheet does not give much information - so it was just found by a more random search and there may well be even better types, especially for small capacitance (special low loss cap series). With low loss factors the DS numbers seem to be upper limits only - hardly any typical curves. This may be in part that there could be a limit to the measurement with normal instruments.

In  a note from Kermet I found the hint that they have lowest DA with CH dielectric - so it may be worth checking one of those types too. C0G is tuned to have a low TC (which is not needed here) so chances are that some other dielectric can be a more clean high isolation ceramic. From the TC CH would still be as good as the film caps.

The absolute capacitance value is not critical so no need for low tolerance.

I would suggest testing something like
C3216CH2E103J115AA and C3216C0G2E103J115AE
There is still a crazy large number of capacitor type to choose from.

[branadic]

Thanks Kleinstein.
I analysed the pictures posted by MickleT. and pictures I made of my unit. Looks like Advantest used different caps in my unit, maybe cheaper and worse ones? The ones in my unit are potted from the top. Need to disassemble the unit again and check, if I can find a brand or something.
Wouldn't surprise me, if they used cheaper caps in R6581D, while yours TiN is model T.

-branadic-

[Kleinstein]

Here is a quick estimate on how much error to expect from the DA. The effect of the slower part (e.g. more than some 2-5 ms time constant) of the DA should about follow the average integrator voltage (MickleT has measured this). The worst point there is the jump of some 1.2 V near zero. With some 20 nF this corresponds 24 nC in the capacitor. The full scale charge is 10 V/20 K *200 ms = 100 µC.

For a really good PP cap one can expect a hidden charge fraction of some 0.01%, maybe a little more for the 1-200ms time scale.
0.01%*24nC/100 µC = 0.024 ppm. There may be another factor of 2 as the charge is transferred to the zero reading and thus entering twice in the difference. This would be about the size of the jump expected for a really good capacitor. With a not so perfect one, possibly after aging the DA may be higher by up to about a factor of 10.
The observed jump for Brandic's unit looks like about 5 µV or 0.5 ppm of the FS. So this would be the DA error for a more average to more poor PP cap.
The different caps could be just a different brand - could be the price or just availability of really good caps.
With the slow modulation the design is definitely relatively sensitive to DA, and should know at Advantest. I doubt the would skimp on the cap - at least if they can still get good ones. At the time NP0 MLCCs may not have been available in good capacity and quality.

The INL error of TIN's unit looks like trouble. With quite some error at +-1 V the ACAL scale factors between the ranges may be off quite a bit. From the plot script it looks like the curve is upside down from the other curves showing straight line - measured value.
For some odd reason the jump is near +1 V and not near zero. This is way to much offset for the amplifier (would no longer work in 1 V range), so it would be something with the ADC.

[branadic]

Multiple capacitor solutions come into mind:
1. better polypropylene foil capacitors
2. C0G/NP0 capacitors
3. Mica capacitors

Meanwhile I analyzed the first of multiple measurements I made and found a correction equation for R6581D, to improve its INL, as it is not limited by resolution.

-branadic-

[TiN]

With quite some error at +-1 V the ACAL scale factors between the ranges may be off quite a bit.

1V scale does not matter, since all meters in my tests are in manual 10V ranges.
  Lame 6581 does not force range when set manually by GPIB command. Retesting now, my bad.
Same with source, which is locked on 11V range.

[Kleinstein]

Multiple capacitor solutions come into mind:
1. better polypropylene foil capacitors
2. C0G/NP0 capacitors
3. Mica capacitors

Mica caps are good for long time stability and very high frequency, but they have rather poor DA in the lower frequency range.
An alternative material would be more like a PTFE cap and also PS (similar quality to PP).
For the ceramics I would also consider CH type dielectric - these are also available as 10 nF.

Instead of testing the capacitors in the meter, one could also do a test outside. Such a setup is not so complicated. In a parallel thread there is an example:
https://www.eevblog.com/forum/metrology/project-micro-voltmeter-design/msg2992348/#msg2992348

[branadic]

It was Conrad Hoffman who tested several capacitors and gave a brief overview of his results:

http://conradhoffman.com/cap_measurements_100606.html

So I searched, found and ordered some comparable cheap NOS PTFE capacitors, that I will give a try to see if I can measure any difference in INL. No, not one of those audiophoolery capacitors, but good old Ukraine work instead:

https://www.ebay.de/itm/560pF-0-1uF-500V-K72P-6-PTFE-Capacitors-fits-brand-name-Teflon-USSR-NOS/193107860442

I also had a look on several different foil and ceramic caps, but they can be even more expensive compared to those NOS PTFE, so I rejected those for the moment.

-branadic-

Edit: Found this on http://www.audio-consequent.eu/info/inf_baue.htm

Die verlustärmsten Kondensatoren sind mit Glimmer, Teflon, Polystyrol oder Polypropylen aufgebaut. Der s.g. Verlustfaktor D wird auch Dissipation Faktor DF, oder Winkel tangens delta genannt. Er wird angegeben bei 1kHz und 20°C und liegt bei Folienkondensatoren im Bereich von 0.005-0.00001 bei Elkos: >0.3-0.02 bei 120Hz (je kleiner desto besser !). Der Verlustfaktor steigt mit der Frequenz. Der Kehrwert dieser Kenngröße wird als Güte oder Qualitätsfaktor bezeichnet. Manchmal wird der DF auch in Prozent angegeben z.B.: 0.1% = 0.001.

Die Ursachen der Verluste sind neben den Dielektrikumverlusten die aufbaubedingten Fehler: Die gering vorhandene Induktivität (Wickeltechnik), der Widerstand der Anschlußdrähte und deren Kontaktierung (Lötstellen bzw. ferromagnetische Materialien). Die Widerstandsverluste, insbesondere bei Elkos, werden als serieller Verlustwiderstand zu dem Wert ESR (Equivalent Series Resistor) zusammengefasst. Dieser liegt in der Größenordnung von 5-30mOhm. Die Induktivität (Größenordnung 10-200nH) bewirkt, daß ein Kondensator ab einer oberen Grenzfrequenz (Resonanzfrequenz) nicht mehr als solcher arbeitet, sondern als Spule. Die obere Grenzfrequenz kann bei Elkos schon bei 10kHz liegen ! Bei Folienkondensatoren liegt der Wert bei einigen MHz.

Eine weitere, selten im Datenblatt (nur bei MIL) angegebene Verlust-Kenngröße, ist die Dielektrische Absorption DA (in %) (>> Kabel). Sie beschreibt eine Art Gedächtniserscheinung bei Kondensatoren (recovery voltage, memory-effect). Ein geringer DA scheint immer dann vorhanden zu sein, wenn der Tangens-Delta auch gering ist. Geringste Werte liefern PS, PP, Glimmer und Teflon Kondensatoren. Die Größenordnung reicht von ca. 0.01% bei hochwertigen PP-Kondensatoren bis >10% bei Elkos und MPs. Ungepolte Elkos sind etwas besser. Erstaunlicherweise besitzen oft ältere hochgeschätzte Ölpapierkondensatoren einen recht hohen DA-Wert von 2 bis zu 20% !  Tolerierbar ist eigentlich nur ein Wert bis 1%. Glücklicherweise läßt sich der Einfluß eines hohen DA-Wertes durch eine niederohmige Beschaltung stark verringern (z.B. beim Einsatz in Frequenzweichen).

Ein hochwertiger Kondensator wird mit einer Kapazitätstoleranz von 0.5-5% geliefert (Elkos 10-20%).

Dielektrikum  |  Konstante ( etar ) | Verlustfaktor (tangens delta)
Keramik  >50    0.001
PP           2.3   0.001- 0.00005
PS           2.5    0.0001-0.0003
Glimmer   4-8    0.0001-0.0002
Ölpapier   3-4    0.0001-0.0002
Luft          1      < 0.00001

Possible PP caps are http://lcrcapacitors.co.uk/wp-content/uploads/mkp-hr.pdf available at Farnell: 10nF: 9520830
Possible PS caps are http://lcrcapacitors.co.uk/wp-content/uploads/fsr-vm-fsr-ax.pdf available at RS Components: 10nF:  426-3110 1n5F: 116-278

[TiN]

Reran test with more points, longer time, NPLC20 and more meters Vertical scale is +0.25 to -0.25 ppm.
Also used average 3458A as reference, instead of source. So plot data shows deviation of each unit from the (3458-3 + 3458-11 + 3458-2) / 3 value.



Still something is off around zero with fake-8.5d-meter.

RAW data
Python app to plot data, conf-file for it
Python app used to capture data

Columns in DSV:
source - programmed voltage in source, source locked in 11V range.
duta - 3458-3
dutb - 3458-11
dutc - 6581T
dutd - 3458-2
dutx - 1281

[dietert1]

We all know how you love your 3458A's.
Somebody else might say: Ah, the maximum nonlinearity of the 3458A in the diagram seems to be about 0,11 ppm. Add 50 % (since that value entered the average used as reference) gives 0,165 ppm. What is the maximum nonlinearity of the Advantest and the Datron units?

Yes, the Advantest curve looks very strange, but that may be some kind of adjustment/calibration problem. Your DATRON unit exhibits a very regular and easy to correct nonlinearity behavior (gain difference between positive and negative input).

Considering the temperature coefficients of 0,34 ppm/K (3458A typical as of Dr. Frank) all these errors are small, aren't they?

Regards, Dieter

[Kleinstein]

The jump at around 0 is the main feature expected from the capacitors DA.
The INL curves that were shown in early publications (comparison to JJA) also has such a jump, though smaller.
The difference could be due to different DA levels of the caps, maybe due to aging or just random variations in quality.
The size of the jump also makes sense for the expected DA level for PP caps.

I don't think the Advantest DMM has a correction for the jump at 0 as part of the calibration. MickleT has looked at the software in detail and never mentioned this. it would kind of make sense to add such a correction in software though.

It would be more like the different gains for the positive and negative side for the DA1281 could be a calibration issue. However I don't know if it uses different scale factors.

For the local peaks at high voltage there can be noise (especially from the references) involved.

[branadic]

With the remaining time on the last day in the climate chamber I tried to get a raw idea of the t.c. of our 3458A and my R6581D, thus only a ±3K change (ACAL/internal cal @ 23°C --> 20°C --> 23°C --> 26°C --> 23°C).
Note, the t.c. prior to the resistor upgrade was around 1ppm/K, measured against my LTZ1000 reference in a thermal chamber at home, while watching the change in voltage readings due to changing ambient temperature.
So t.c. has significantly improved, though is still twice the value of our 3458A, but infact not to bad.

-branadic-

[ramon]

Branadic, thank you for your good work with the thermal chamber. That's the way to go!
 
Althoug T.C. for R6581 is twice than 3458A, avid observers looking that graph with lupe will note that R6581 provides a flat line two times faster than 3458A.
In fact 3458A is not able to provide a flat line when we step down from 23°C to 20°C and up again to 23°C and 26°C.




This can be a trap for young players (and for metrology experts as well, that calls R6581 a fake 8.5 digit dmm  ;-)

(this reminded me this topic ...

https://www.eevblog.com/forum/metrology/t-c-measurements-on-precision-resistors/msg1169432/#msg1169432

"try a test where you hold temp for a dwell time to find out the thermal time constant of your test setup.
In other words, change temp 5 degrees, then hold temp steady until you see resistor stabilize, then
change again and wait until you see the resistors really stabilize, etc.
Until you do that test you won't know how fast your test setup can respond to temp changes.
You might find out your temp. ramp rate has to be much slower. 
We don't use a constant temp ramp, we've found that a dwell time at each temp level is a little safer.
That will show you if you're outpacing the resistor's ability to stabilize to the new temperature.

https://www.eevblog.com/forum/metrology/t-c-measurements-on-precision-resistors/msg1170432/#msg1170432
https://www.eevblog.com/forum/metrology/t-c-measurements-on-precision-resistors/msg1170456/#msg1170456
https://www.eevblog.com/forum/metrology/t-c-measurements-on-precision-resistors/msg1170462/#msg1170462

"In other words:  Change temp, and don't change temp again until your resistance measure line gets flat again.  It can take a while." )

[TiN]

Somebody else might say: Ah, the maximum nonlinearity of the 3458A in the diagram seems to be about 0,11 ppm. Add 50 % (since that value entered the average used as reference) gives 0,165 ppm.

I do love my 3458s. I didn't quote understand where you get 50% from and why we need to add it? Care to elaborate a bit more?

Completed run on 5720-2 with exactly same setup/meters/settings.
RAW plot as is from instruments. Interesting to note absence of 0V shift we seen before from 5720-1. Otherwise dataplot looks very similar and difference between each DMM is same.



Same plot against 5720-1 for reference, captured earlier with same settings/app:



Also here is RAW data from second 5720A.

Obviously source INL error is removed when I've used 3458A as a "reference INL" value. Here is it seen as flat bright cyan line. So next plot shows deviation of each individual meter against average of three 3458A:
Yes, far from ideal method to test INL, but I don't have better idea how to get source with better than 0.1ppm INL, other than JVS system.



Same data for 5720-1 source (same plot posted earlier):



P.S. Uncertainty is the reason why all these graphs called "system linearity" and not the "DMM linearity". There is no way to figure out what is individual INL error of either particular DMM or source, without making whole range of assumptions like "we guess that INL of three 3458A's is still inside 0.1ppm spec" or "<0.5C temperature deviations in lab don't impact data much". So at best we have approximated maximum error of the whole tandem.

[Kleinstein]

The factor 1.5 is needed as the 3458 under test is also used in the reference. So for one of the 3458 the difference is meter A - (meter A + meter B + meter C) / 3 = 2/3 A - 1/3(B+C). So one could argue if the right factor is 1.5 or maybe  1/sqrt( (2/3)²+(1/3)²).

The other problem in judging the 3458 from just comparing similar meters is that the INL error can also be correlated. The comparison of same type meters can hide some systematic errors. If done with 2 of the 6581, the just near zero may appear much smaller, because it seem to be a common feature.

It looks like the average of the three 3458 is about the best one can do as a reference. One could still combine the runs with both sources. This would reduce the noise and some drift effects.

The calibrators both seem to show some periodic linearity error. In addition there seem to be some offset causing the jump around zero.

[dietert1]

In the end you can't define a linear reference by democracy, whether you take the votes of all or only the elite or what you think is the elite. Just imagine someone had three Advantest DVMs and used them to test a 3458a. So that approach isn't very reliable.

As far as i understand a stable divider combined with some mathematics can give you a valid answer, independent of what other equipment you have around. For example:
https://www.eevblog.com/forum/testgear/dmm-linearity/?action=dlattach;attach=157863

Regards, Dieter

[Kleinstein]

Those concepts with a divider or chain of reference source can give a few tests of self consistency. Usually the number of points to test is limited to a relatively small number. The more steps are in involved to larger the certainties from noise, thermal EMF, EMI and input currents. So it may need relatively low value resistors or extra buffers. Sufficiently stable dividers is not an easy task.

The two methods are kind of complementary.  A few points from stacking voltages and the comparison with other instruments in between. The 3458 meters are known to be usually good for better than 0.1 ppm, so they are Ok have a look at errors in the 0.2 ppm range for the 6581T. To get more confidence in the comparison one can use a kind of dithering, adding some offsets (e.g. 1 V range)  between the meters.