Prema 6048 teardown

[dietert1]

Your nonlinearity curve near zero looks like what branadic was showing: The plus and minus range don't meet ends at zero, with an offset of about 24 uV in your case. The shape of the curve indicates that the instrument does not remain in one of the ranges while within a +/- 100 mV or +/- 40 mV margin, but it is using both ranges.
How is that possible? To start with, the instrument does use both ranges for zeroing. When you get close to zero the instrument can use the sign of one measurement to determine the + or - range for the next measurement. So it preserves the +/- margin around zero for some AC in the input signal. If one takes averages, very close to zero one will have the observed nonlinearity shape with an edge as sharp as noise allows.
The 20 uV to 24 uV offset appears very similar for several instruments - likely a constant (error) built into the firmware. Dirt effects like thermal EMF of the polarity relay might add a uV or two.

Regards, Dieter

[Kleinstein]

I don't think the curved from between some -100 mV and +100 mV is due to averaging over using the + and - relay setting. Most of the data shown here are with a rather long integration time of some 20s / 40s and thus not averaging very much. The jumps would also be visible and I would not expect a relatively smooth curve.  It is more that the meter stays with the old sign until it reaches some -40 mV and than changes the sign. The switching takes time for settling and thus wants some hysteresis to avoid overly frequent switching.
One of the older curves from brandic showed the contineous curve down to -40 mV and only than a jump.
When using the KVD one may have forced the switch over already near 0 mV, e.g. from the sequence the points are measured.

My best guess on the nonlinearity is from the ADC getting nonlinar when getting close to the low limit of the hardware (some -120 mV theoretical, maybe -100 mV if pulse skipping is excluded). At the very edge the pulses for the reference are short (like < 1.5 µs for 120 mV from the edge). Due to noise and hum some pulses may end up even shorter (the clock is not very high), and depending on the HW possibly are even missing.  It is expected to get some nonlinear response very close to the edge - the question is only if the 120 mV distance to the edge is sufficient. I understand that they want to keep it small, as there is some drift with the shift.
To me pulses of some 1.5 µs look rather short for an not that fast (typ 600 kHz GBW) OP77 in the integrator.

I don't think the nonlinear part is from some numerical correction that is going wrong. The more logical countermeasure would be a slightly larger shift.  A numerical correction could be tricky, as I would expect hum to change the INL (make it worse). It would however need quite some hum (e.g. > 20 mV) to make a difference.

When assuming an ideal curve through zero, the data would be slightly out of spec (~0.6 ppm off). If one could have some residual offset for a short, one could get parameters to reduce the INL error to maybe 0.3 ppm and thus in spec.  It would still be rather unconventient to have an offset and it would also need extra effort to set it that way.

[Rax]

To me pulses of some 1.5 µs look rather short for an not that fast (typ 600 kHz GBW) OP77 in the integrator.

I think it should be a pretty easy test to drop an OPA277 in there and see where linearity lands. I assume calibration would be thrown off (but undisturbed when the old op amp is being swapped back - thoughts?), but I think the plot going through zero should be visible.

[Kleinstein]

Yes, changing the OP-amp would effect the calibration: the offset of the OP-amps is subtraceted from the 7 V reference.  Changing back to the old OP-amp would largely bring back the old offset and calibration.
One should still first to a comparison to avialable other reference before the test.

[branadic]

Nice to see, that my observations were confirmed independently, so it's neither a fluke resulting from my measurements nor something specific for the meters on my desk.

Yes, changing the OP-amp would effect the calibration: the offset of the OP-amps is subtraceted from the 7 V reference.  Changing back to the old OP-amp would largely bring back the old offset and calibration.
One should still first to a comparison to avialable other reference before the test.

I have an OPA202 coming along and can give that a try too. As far as I understand no adjustment is required for the investigation process, it's enough to see if the shape of the INL curve has changed or not. So pretty much straight forward.
I was also in contact with someone else having a white 6048, which was part of an INL-testing at Prema as he said. At least in the 2 V range its INL was <0.2 ppm. I'm still waiting for number of the higher ranges though.
The so far seen ±0.5 ppm INL in the 20 V range is not too bad, but not exceptionally good either, my Solatron 7081 shows ±0.3 ppm, still way larger than a typical 3458A and thus limiting the instrument for voltage transfers.

-branadic-

[Rax]

Nice to see, that my observations were confirmed independently, so it's neither a fluke resulting from my measurements nor something specific for the meters on my desk.
-branadic-

I'm happy to contribute any useful empirical data that is relevant to the assessment of this unit.

One aspect where we may not have the advantage of a completely randomized set of lab data inputs (between my bench and yours) is the fact that, as far as I know, your source (AN3200, correct?) and mine (DP8200) are AFAIK essentially the same unit. I've collected data with the FX while I still had it here, and will seek to generate the INL plot with those too. Also, the fact that I used a KVD (not sure what your methodology has been, branadic) may introduce a bit more randomization of the test conditions, which would contribute positively to the relevance of the results. For instance, I assume both the linearity and the offset at zero of the DP8200 are not a factor.

But I'd just make sure the offset on your AN3200 is at a reasonably low value (the SM stipulates under 10uV). Not entirely sure if a factor for this (depending on your test setup) but may play a role.

[branadic]

As you may haven't noticed, but on the x-axis I have the reading coming from a Solatron 7081, not what you call stimulus. That means, whatever the AN3200 is throwing out in reallity doesn't influence the x-axis.
However, the AN3200 was fully maintained and adjusted against a 3458A.

-branadic-

[Rax]

As you may haven't noticed, but on the x-axis I have the reading coming from a Solatron 7081, not what you call stimulus. That means, whatever the AN3200 is throwing out in reallity doesn't influence the x-axis.
However, the AN3200 was fully maintained and adjusted against a 3458A.

-branadic-

I did read your graphs, branadic, though I can't say I have clarity on your test setup and data collection means. I assume there is a stimulus involved (I think you stated that to be an AN3200?), though I also assume you're collecting the x axis data with your Solartron? Such that the x would become sort of source-agnostic?  But please explain this in as much detail as you're willing (we can take it offline if you prefer, but I'd appreciate getting a good understanding).

From that perspective, I'm not sure I understand how you're making sure the x data is at its turn agnostic of the Solartron's own non-linearity.

But my comment was more general, in due diligence relative to our test conditions. I think disclosures are important to make this work useful to others.

[Kleinstein]

Much of the nonlinear effect seen near zero is in a rather limited range  (e.g. +-200 mV). So one does not really need a high grade instrument to compare too. If really needed a HP34401 in the 100 mV or 1 V range would be good enough to see the deviation of some 10 µV.
Unless there is a problem with the calibrator, the AN3200 or DP8200 should be good enough for the tests too.
It would need a good instrument of other methods / tricks to check the larger range for possible additional more soft INL errors, like a U² or U³ contribution. So for they are not visible. This part would be one where the quality of the source or reference meter can be an issue.

[Rax]

Much of the nonlinear effect seen near zero is in a rather limited range  (e.g. +-200 mV). So one does not really need a high grade instrument to compare too. If really needed a HP34401 in the 100 mV or 1 V range would be good enough to see the deviation of some 10 µV.

So, in effect, the core argument there is the enhanced granularity and accuracy of a lower range of measurement of the instrument used as a reference. I get that. Even in that case, though, the errors would still be cumulative, correct?

Unless there is a problem with the calibrator, the AN3200 or DP8200 should be good enough for the tests too.

I'm not so sure I follow you here, though. Those calibrators are +-10ppm of rdg + 10uV on the 10V range (the 100mV range need not apply for related and other reasons). Getting the offset to be much less than 10uV is a challenge too (probably explains the formula above). This is why I'm using the KVD which is a .1ppm accuracy instrument. I just don't see how those calibrators would cut it, but maybe I'm not focusing on the right criteria.

[branadic]

There are two ways of measuring the INL:

1. Use a well known, highly linear source
2. Use "any" source and compare it to a well known linear meter (such as 3458A or similar)

I do the latter, as I don't have such highly linear source. But as you already have proven using methode 1. we both come to the same conclusion, that is, there is a linearity issue with the Prema 6048 at least in the 20 V range.

-branadic-

[branadic]

So OPA202 arrived yesterday. I gave the unit one night after I've replaced OP77 with OPA202, zero'ed it and measured INL again. Attached the file of the measurement as well as a diagramm with the outcome of it.

-branadic-

[Rax]

So OPA202 arrived yesterday. I gave the unit one night after I've replaced OP77 with OPA202, zero'ed it and measured INL again. Attached the file of the measurement as well as a diagramm with the outcome of it.

-branadic-

I appreciate the work put in this to test it, branadic. Do I see this right - does it seem about maybe a bit over half the spread of what the stock op amp was giving out? So a path worth taking, an improvement?...

[Rax]

I'd probably also like to see an assessment of the impact of this on the state of accuracy of the instrument. In case the INL performance improvement throws the baby with the water...

(I may also be able to perform this assessment, having an OPA277P in my supplier's cart... Not sooner than next weekend though.)

Of course, one could always re-"software adjust" the instrument given some artifact capabilities, but if that's not available, this is an aspect to consider.

[branadic]

I wouldn't think that the OpAmp made an improvement, it's more like the process of penetrating the unit made the difference. If you followed the thread you have seen that the unit which I call 6048b changed it's INL behavior during the process (paralling resistors) and even changed polarity, although I brought it back to its original settings (removed parallel resistors), only interchanged both OP07 within the unit.

1. https://www.eevblog.com/forum/metrology/prema-6048-teardown/msg4657387/#msg4657387
2. https://www.eevblog.com/forum/metrology/prema-6048-teardown/msg4659679/#msg4659679
3. https://www.eevblog.com/forum/metrology/prema-6048-teardown/msg4668043/#msg4668043

-branadic-

[Kleinstein]

I'd probably also like to see an assessment of the impact of this on the state of accuracy of the instrument. In case the INL performance improvement throws the baby with the water...

(I may also be able to perform this assessment, having an OPA277P in my supplier's cart... Not sooner than next weekend though.)

Of course, one could always re-"software adjust" the instrument given some artifact capabilities, but if that's not available, this is an aspect to consider.

There are a few side effects with the change of the OP-amp: the OP-amp at the integrator effects the calibration (scale factor - offset relative to 7 V and thus some 10 ppm of possible change).  Offset drift of the amplifier would also add to the long term drift. The old OP77 in a ceramic case is likely quite good in that respect. A plastic case may add some sensitivity to humidity.

A software correction is an option, as the INL part seems to be quite stable. This would likely be correction of the readings, likely in an extra step on the PC side. It looks a bit too complicated for the integrated math functions.

The curve with the OPA202 looks a little better, though still not really good. Part could be from the offset adjustment. It would need a closer look at the data closer to zero to really tell. Even if only a partial fix, it supports the amplifier as a possible weak point and it could be worth trying a different amplifier. An AZ OP-amp (e.g. LTC2057 with adapter) could be somewhat tempting, though the ones with a +-12 V supply range tend to be a bit on the fast side. So there would be a chance for instability (oscillation), but also the chance for good performance (e.g. reduced noise and INL and drift) if the integrator stays stable. Already the OPA202 should have slightly lower noise than the OP77 (but likely not much - the very low frequency data are a bit uncertain).

[branadic]

The curve with the OPA202 looks a little better, though still not really good. Part could be from the offset adjustment. It would need a closer look at the data closer to zero to really tell. Even if only a partial fix, it supports the amplifier as a possible weak point and it could be worth trying a different amplifier. An AZ OP-amp (e.g. LTC2057 with adapter) could be somewhat tempting, though the ones with a +-12 V supply range tend to be a bit on the fast side. So there would be a chance for instability (oscillation), but also the chance for good performance (e.g. reduced noise and INL and drift) if the integrator stays stable. Already the OPA202 should have slightly lower noise than the OP77 (but likely not much - the very low frequency data are a bit uncertain).

I'm still not convienced about that. As I stated above, the unit that I call 6048b changed INL behavior and "polarity" during installing and removing parallel resistors and for some odd reason the INL improved (dropped by a factor of 2) along the process, but still having the very same OP77 in ceramic package installed. How do you explain that, magic healing of OP77? 
Unlikely. I wish we could have a more systematic step-by-step approach to figure out what's going on.

-branadic-

[Kleinstein]

The curve with the 6048b showing the opposite sign is missing the low voltage points. The 0 V reading in the curve looks like it is quite a bit off (-25 µV). So chances are that the opposite sign was just due to a missing or somehow wrong (effected by noise or thermal EMF or EMI) zero adjustment. Though only 1 point in the curve the reading at 0 V suggests a problem with the zero adjustment. The 0 V point is even about similar 10 µV lower than the extra polation from the positive values - so just like the later bahavior and the other meters.

So far the curve with the OPA202 only has points at 0 V , +-50 mV and +-100 mV.  Here the +-100 mV still look like they fit to the higher votlage part. So it is a bit hard to see if and how much the nonlinear part has moved to lower voltages. For the original HW there is a curve with 10 mV steps. I may make sense to also get similar data, e.g. from some 100 mV down to -40 mV. This would allow for a more direct comparison.

The OPA202 is only somewhat faster (1 MHz typ compared to 600 kHz typ). From the limited data so far and the old curve it looks like there is a comparable error for the OPA202 at 0 V and the OP77 at some 50 mV. This would be distance to the theoretical ADC range of 120 mV and 170 mV. So some 30% reduction in the settling time needed.  There was some hope for more improvement, but the GBW is no the only factor to determine the settling time.  I still consider some 40% reduction in the INL error encouraging.


For the HW side some part of the integrator is inside the ADC chip (pins 11 and 13). We don' know what is hidden there - it could be just a resistor, a fast buffer or maybe some amplifer.
With no extra access to the summing node it can't be the classical 2 OP-amp integrator or the circuit as the Prema 5000 and similar.
So it is unclear how fast the external OP amp for the integrator can be.  Similar it is hard to simulate the settling with part the circuit unknown.

[Rax]

I wish we could have a more systematic step-by-step approach to figure out what's going on.
-branadic-

That's why I am so hung on methodology and approach (I am a physicist by background, so my mind is hardwired this way). I am very satisfied we've followed two independent methodologies and arrived at the same result. That's ideal.

[Rax]

Removed the parallel resistor to R36, shuffeled of few IC's around, gave the unit a good warmup, then performed nulling on all three units in auto ranging, adjusted 10 V range to the same 10V reference and performed INL test again.
For some odd reason the INL of the P6048b is now smaller than the others, but still visible, while on the other two units it hasn't changed at all.
So I guess the only thing left we can do is to analyze, understand and improve firmware for the 65C02 processor. Martin already started with that.
-branadic-

I missed this message, it looks like the improvement was already present at this point? I compared the OPA202 readings with the earlier, initial ones. Maybe the magic sauce is just conditioning the legs of the ICs and reseating?

[Rax]

Well, that's interesting. It seems I just lost C61 - I thought there's an odd smell at the bench, but I have no idea when this happened. I've seen dipped tantalum caps fail and it's not pretty. I'm thinking low-esr, 8-10,000hrs service electrolytic - I don't typically replace tants with fresh tants. Any reason to go a different route here?

[Kleinstein]

C61 is at the supplies, between an 7815 and 7812. A low ESR Al electrolytic should be OK. I see not need to have a tantalum, unless one wants to operate the DMM well below freezing temperature.

[Rax]

C61 looks pretty toasty too, I typically use a WIMA film for something like this. And on further investigation R1 and R2 seem to need to be replaced. A little PS project on my hand. 

[bdunham7]

Any reason to go a different route here?

No, not in an LM78xx circuit. It was probably a lower voltage version?  You could use 35 or 50V tantalums, but Al El-caps will work just fine here and you likely don't need anything too special.  $0.14/ea should cover it.

https://www.mouser.com/ProductDetail/Wurth-Elektronik/860240572001?qs=sGAEpiMZZMvwFf0viD3Y3aZipiehufnXaZqtKrG5UnKry1vdQEzUvA%3D%3D

Has its neighbor C4 already been replaced?  I'd look at C62, C4 and C5 as well just to be thorough.

[Rax]

Has its neighbor C4 already been replaced?  I'd look at C62, C4 and C5 as well just to be thorough.

Nope, C4 is still there (the only two tantalums in sight - why this choice?...), but I am now thinking this needs a full recap (of the PS). I'd typically bump the values up to twice the value for the larger ones (220 -> 470, etc.), though stay with value for the lower (<=100uF or so). I typically use Nichicon HE, PW, or other rugged, low-esr, long life (>5k hrs, but more like 8-10k) caps in instrumentation. I also like Panasonic (FR, FM) and Elna for some extremely low-esr (as measured @ 150kHz) caps, but typically Nichicon offers longer life at my regular supplier.