Prema 6048 teardown

[branadic]

Thanks Mickle T., I've quickly recreated the measurement of David Partridge aka perdrix:

https://www.febo.com/pipermail/volt-nuts/2018-January/005593.html

and performed that test with the units set back to its factory constants. Indeed it looks identical, but on the other hand we have never heard if that firmware update fixed that issue. 
Seems like I have to dump and reprogramm the EPROM with "new" firmware but original factory calibration constants and check INL afterwards again, first with factory settings and with adjusted range.

-branadic-

[branadic]

I've flashed an OTP with the latest firmware and the factory cal data and performed the quick INL test in the 20 V range on the same unit again. While the offset issue was solved, the INL picture hasn't changed at all.
From my quick phone call with Systems Engineering Kalibrierlaboratorium GmbH & Co. KG, the cal lab that performed the DAKKS cal for Prema back in the old days, I've learned that this INL is to be expected. I'm still waiting for a response to the INL plot I've showed them, but I'm not expecting surprises.

-branadic-

[Kleinstein]

The INL curve really looks like an offset (some 11 µV) for the ADC itself that is not included in the right way. The actual offset can be larger, the 11 µV is the part not handeld right.
The ADC offset may drift a little over time and it would thus make sense to include correction of the ADC offset with the zero function (with external short).  Is the zero function switching the polarity relay to measure the zero for both polarities  ?

The only strange point is the zero readung in the INL curve. I would have expected a different zero readings depending on the polarity relay setting. A zero may come up by averaging both cases (coming from pos or neg values).


For the linearity clams on some meters there could be a confusion with the turn over error. Some meters note the turn over error as a measure of linearity, though this is only 1 point - maybe the easiest to measure.

[branadic]

Maybe I'm getting you wrong or I simply don't explain everything detailed enough.

1. Performed zeroing in 20 V range
2. Adjusted 20 V range with known 10 V source
3. Performed INL measurement in 1 V steps
4. Calculated deviation from straight line (first order line fit) taking all measurement points between -10 V and + 10 V into account

Looking at the result the zero point is almost near zero, but the +/-10 V point are 10 µV too high/low. Wouldn't it make more sense to use the end-point INL instead (0 V and 10 V points), as these two were adusted before?
The result would be, that 0 V is at zero deviation, 10 V is at zero deviation too and in between we see some negative deviation for the positive part and a positive deviation for the negative part, while overall INL would be way lower.
We also have to remember that 10 µV is 0.5 ppm of range for this instrument.

-branadic-

[alm]

From my quick phone call with Systems Engineering Kalibrierlaboratorium GmbH & Co. KG, the cal lab that performed the DAKKS cal for Prema back in the old days, I've learned that this INL is to be expected. I'm still waiting for a response to the INL plot I've showed them, but I'm not expecting surprises.

With the adjustment procedure that only adjusts a single point there can't be a linearity adjustment. So unless the meter had an autocal-like mechanism for linearity adjustment, I wouldn't expect INL to change at all after adjustment. There may or may not be a factory adjustment for linearity, but I don't think we've seen any evidence for that.

[Kleinstein]

The INL curve shows a jump of some 22 µV in the INL error from -1 V to 1 V, just in the area of polarity reversal. The 0 V reading is a bit odd - it could still be the average of reading in the positive and negative range. No knowing the details I leave that point out for now.
Except for the jump at the polarity reversal the rest of the curve is pretty linear. So if one would correct the data with an -11 µV shift for positive and + 11 µV for the negative relay setting one could get a much better INL curve. The residual linear part would be removed from new line fit or end point curve.

There is only a single external voltage for the zero adjustment, but there should be still 2 zero readings done for both settings of the polarity relay. The difference of the two gives the offset of the ADC and the average gives the offset of the input amplifier. So a correct procedure for the zero could correct for the ADC offset and the amplifier offset and thus just for the type of problem seen here.

There could of cause also be a linearity problem at the ADC, for readings very close to 0 V, so that the zero adjustment does not work correctly and the error gets visible for the larger voltages.
It may help to have another look at the INL curve for a few more points between -1 and 1 V , making sure to keep the data for both relay settings separate.


Just in case: it might make a difference if the zero adjust is done in auto mode for all ranges or just for one range with manual ranging (this could accidently limit the zero adjust also to 1 polarity - which would be bad).

[branadic]

The zero adjustment was performed only for the current range of interest, I didn't use zeroing in auto ranging mode.

During adjustment of the 20 V range I can confirm the INL error already. I have a well known 1 V / 10 V reference. If I use the 1 V (5% of the 20 V range) to adjust the 20 V range, then apply 10 V to it, I can observe a deviation of the 10 V reading. If I use the 10 V reference for adjusting the range and apply the 1 V afterwards I can see the very same amount of deviation too.

Looking at the datasheet of ADC5601 it claims 0.0001% INL so 1ppm. It further mentions "Gain and offset adjustments are normally made by the microprocessor which controls the analog to digital converter...". Assuming a similar behavior of the BK7 ADC I can only draw the conclusion, I can only draw the conclusion, that there is a bug in the firmware, leading to the observed results.

-branadic-

[Mickle T.]

Assuming a similar behavior of the BK7...

I think BK7 == PRI5610. The last have up to 0.08 ppm INL (PRI5610E).

[Kleinstein]

I bug in the firmware is possible explaination. However I wonder why such a rather visible problem (quite some discontinuity when the polarity relay switches and likely 0 V not reading as 0 in one of the settings) would not have been recognized before. This would be a rather gross fail, not just cosmetics.  Of cause the extend of the error can depend on the individual hardware and with new meters the error could have been smaller or a hidden factory cal values to take care of it in the initial state.  Getting the same INL curve with a new ROM and thus likely different factory cal values sugests there is no such value however.

The 0 V point would be rather close to one end of the ADC range and that can be a bit more prone to trouble (e.g. supply ripple or hum reaching to saturation at times). With internal offsets the ADC may not work well all the way to 0.

One could try to intentionally change the ADC offset and this way see, if there is at least some compensation for it. Allowing for a little more overlap may also reduce the chance for reaching the limits too early.

I do not fully understand the ADC circuit, but it looks like the part around U11,R34,R35,R36 should add a little offset to the input signal to the ADC to get some overlap and zero in both ranges.  A resistor (e.g. 100 K range) in parallel to R36 would increase that shift. Ideally the zero adjustment should take care of such a shift - if not it would cause a huge INL error / discontinuity (e.g. some 100 x the current).
A large error would kind of proof that it is a software side (may still be cal constant) problem. If it is a range limit problem of the ADC things may get a little better, though possibly not yet perfect with a relatively small shift in the 5 mV range.

[branadic]

Mickle T., the PRI5610 could be a version of the BK7, yes, but both have different package with different pin quantity and we don't know for sure as there is no datasheet for the BK7. The patents don't show more details either.

I guess you mix up a few points Kleinstein.

1. The latest firmware I flashed into the OTP has the original factory cal constants copied to it that I restored from the original EPROM before.
2. There could be a small offset at 0 V coming from the source (Analogics AN3200) that has to be ignored and is not the problem as I compare the Prema 6048 against the Solatron 7081, not against the set value at the source.
3. When I short the input the Prema 6048 reads zero, so again, don't worry about the offset.
4. The large INL at 1 V, falling of towards 10 V, is real and not just a coincidence, but I'm performing a smaller step INL measurement right about now, will see how that turns out. I also increased soak time and take a few more samples at each voltage.

-brandic-

[perdrix]

Maybe the attached firmware that Prema sent me will help somewhat

David

[branadic]

That is the firmware that I'm currently running my INL tests with, thanks.

Attached is a measurement with more details between -1 V / +1 V. As can be seen it's not an offset and how it's handled internally, but is a real INL effect of the meter.

-branadic-

[Mickle T.]

Wow! There are over 800 points on this graph. This would require over 11 hours of measurements on the Solartron 7081.
branadic, can you share CSV data?

[Kleinstein]

The INL curve shows a ratgher sudden jump in the INL error around 0 V. The other ranges, when looked at separately do not look that bad or nonlinear. Depending on the assumed gain curve (end point, least square fit, nominal) the INL curve would get a straight line added or subtraceted. The problem in the INL is the jump near 0 V, not that about straight sections for the negative and positive sides. The residual slope is a gain error, that would go away when the jump near zero would be gone.

The problem is not the offset that is visible to the outside or the offset of the source for the INL test. The point is than internaly with 6048 the firmware has to handle 2 separate offset values for the input amplifier and the ADC itself. One could also see this as seprate offsets for the positive and negative polarity settings. If that is not done correct one gets twice the error in the ADC offset as a jump when the polarity relay switches. Chances are the firmware already does some correction for the ADC offset, as the error is only 10 µV. Beside a software problem this could also be an issue with the ADC itself, so a HW point. The 0 V reading is quite close to the edge of the range of an unipolar ADC and there is a chance for an extra INL error of the ADC when very close to the edge. In this case it could help to add more offset to the ADC (the R34-R36 part) to get more overlap and shift the weak spot of the ADC outside of the used range. 

If it is a software problem, I would expect an externally visble offset for one of the relay settings. So the 0 V input one may get 2 different readings depending on the polarity relay setting.

[branadic]

I have added the file to the previous post.
I will start a similar measurement with all three Prema 6048 in parallel soon to proove that it's a systematic issue, not a unit specific issue.

-branadic-

[dietert1]

I am wondering what this meter will do with near zero DC yet with some low frequency AC superimposed. Will it try to operate the polarity switch following the AC?

Regards, Dieter

[branadic]

Meanwhile a fast INL sweep on three units in parallel was performed. No, they are not all mine, one is here for service - a white unit (red dots) - and one for repair (green dots), which showed some additional issues that I try to fix by comparing it to one of the other units. However, the unit that needed repair shows the worse INL with oposite direction, while my unit though the oldest among them all is slightly better to the white unit, but not by much. Overall they are both identical. So this proves a rather systematic behavior.
If I find the difference between the units we can probably find a way to improve INL.

Beside a software problem this could also be an issue with the ADC itself, so a HW point. The 0 V reading is quite close to the edge of the range of an unipolar ADC and there is a chance for an extra INL error of the ADC when very close to the edge. In this case it could help to add more offset to the ADC (the R34-R36 part) to get more overlap and shift the weak spot of the ADC outside of the used range. 

I can't see how that adds an offset to the ADC, can you elaborate on that?
The schematic is barely readable, so the connections are hard to spot. However, R34-R36 are wirewound 1% resistors, R31 is hermatically sealed. It would be possible is to use a 10k resistor network such as TDP16031002 to create R31 (5k), R35 (10k) and R36 (40k) from the very same device. R34 has a value of 750k, nothing too common as precision resistor.

-branadic-

[alm]

I think the new sweeps are really interesting. Based on that, I'd say that:

• This behavior is clearly indicating a design issue (software or hardware).
• Based on the green series the readings near 0 V can be pretty much anywhere. They don't have to be in between the -500mV and +500mV values. Was the green unit zeroed before the sweep?
• The slope (gain error?) can have either sign. This is consistent with the theory that this is an effect that is partially but not fully corrected.

[Kleinstein]

The U11 and R35/R36 part is creating a negative reference of 1/4 the main reference.
This voltage is than used to add some negative current via R34 to the summing node where the input signal via R29 and the switchable refrence current via R31 meet. From the sign it is opposite of the reference and thus allow getting a little beyound zero. 

The part with U10, the 1M resistors and DG308 CMOS switch chip acts as a precission switch with sense from the output side.So there is either current flow of some 1.5 mA or not.
I don't any other connection of the reference to the ADC chip - so chances are high the DG308 would be the main ref. switching of the ADC.

The gain setting resistors to set the ADC gain in relation to R31 should be R31/R32/R33 from the amplfier PCB/part. They somehow use the same resistor lables there, which can be really confusing.

I don't think it would make sense to replace R31 as it has to be matched to the 3 other ones. One would need a matched set of 10 K , 2*5 K and 500 ohm.
Because of the rather small contribution R34,R35,R36 are much less critical - they could still have drifted over time and caused the problem (a factory cal value not longer good). Ideally I would keep the 750 K and if required change the R35/R36 pair for the voltage gain. Matching from R31 to R34 would be nice, but no need to also match the R35/36 pair. The the start as a test it would likely be soemthing in parallel to R36 to see if there is change at all.
The current added through R34 (750 K) is rather small:  1/600 the reference or 1/60 of the 20 V (+ overrange) range. So this should be only some 0.3-0.5 V in the 20 V range and quite a bit of this would be lost to a minimum time for the ref. Switch to be on. To me it already looks on the low side to start with, with very little overlap.

A combination of low DC and AC could be a problem for the DMM. Too much hum (in the signal) or maybe the supply could than cause linearity problems for a voltage near zero.

[branadic]

I have used the P6048b which showed the opposite INL with respect to the other two and attached a 5M resistor to R36. As can be seen the INL curve has flipped and is slightly smaller in max value.
I already prepared a TDP1603 network yesterday to fit instead of R35/R36 to see if that makes any difference.

-branadic-

[Kleinstein]

The change in the jump is from -35 µV to +24 µV or some 59 µV of change.
The shift in ADC input zero should be ref Voltage(~7V) * 10K/5M * 50K/750K = 930 µV for the 20 V range. If not corrected for this shift should contribute twice to the jump near zero.
The 50 K /750K  factor is R33(amplifier board 20 V range) /R34.

So the shift in the ADC zero point is corrected to some degree (some 98.5%), but for some reason not perfectly. Somewhat doubt it, but it could be a serious INL problem of the ADC for low voltages (like < 100 mV), so that the zero correction is that far off.
Another possible cause could be a software problem, like missing a calibration factor (or using a wrong one) for the ADC gain in the zero term. Some 1.5% off from nominal gain is plausible (e.g. for the 7 V ref., maybe a resistor ratio)

I don't think one would need a resistor network for the R34/R35 couple. I see no reason why an accurate 10:40 ratio should solve the problem. 
The 5 M parallel to 40 K is a bit less than 1% change, so quite a lot compared to resistor stability. So there is no need for extra high stabilty, but more for just the right ratio.
Trimming the R34/R35 ratio is not ideal, but could be a relatively simple  fix for the INL. It would still take quite some time for 2 x INL measurement with different settings and use of the ouput of U11 as an intermediate reading, assuming that the jump is linear depending on that voltage.

[branadic]

I used what I had at hand, above 820k the choise in my stock is pretty small. So I've used 5M||40k.
The reason for my idea to use a network for the R35=10k and R36=40k is good thermal matching, but also the ease of adjustment by adjusting individual 10k elements with lower parallel resistors instead of something in the mega ohm range. In this configuration 4x 10k elements can be used to form the 10k and the other 4x elements to form the 40k, with the elements in the network used in an interdigital fashion.

-branadic-

[branadic]

While I can flip the sign of the INL curve with a parallel resistor to R36 as shown before, adjusting it's value doesn't change anything and INL stays the same.

-branadic-

[Kleinstein]

While I can flip the sign of the INL curve with a parallel resistor to R36 as shown before, adjusting it's value doesn't change anything and INL stays the same.

-branadic-

So not a simple factor that is wrong with handling the ADC offset.

If the error stays the same or just 2 possible scenarios (signs of the INL error) a possible point could be having the Zero adjustment done only with 1 relay setting (as a failure in the Zero procedure) and the initial offset determines which setting is used. Is the polarity relay actually switching during the zeroing call ?

A point worth trying could be doing the zero procedure when in autoscale mode and thus doing it for all ranges together. Ideally it should not make a difference, but one does not know until one tries.
The full adjustment may do things a little different (maybe just waiting time) than doing the zero for only 1 range at a time.

A possible source of error could also be leakage between the 2 supplies for the ADC and amplifier part. Especially the current source for the resistance could be an issue as the PM relays may show some leakage - not sure if K13/K14 get a clock/drive all the time, or only when a resistance measurement is active. Ideally only the pins 9,10 and 11 of J4 should provide a link. K13/K14 should switch alternating with enough pause in between so that not much current flows there. With meter off and the connector J4 removed one could check the isolation across K13+K14. These should be pretty low leakage (e.g. 200 pA to cause a 10 µV error).

[branadic]

Checked that. In manual range it seems to run through the same nulling twice without switching the relay, while in auto range it changes polarity in each range for nulling. So some odd behavior indicating a firmware bug?

-branadic-