Keithley 2700 rate change affects the measured voltage?

[petemate]

Hi guys, this is another question in my looong quest to fix a Keithley 2700. The meter is running, but I am seeing a strange issue: The read voltage changes significantly depending on the sample rate(slow, medium, fast). For instance, I am putting in a 10V, but my 2700 shows the following:

Fast rate: 9.966xx
Med rate: 9.972xx
Slow rate: 9.972xx

xx represents numbers that are changing(e.g. due to noise), while the other numbers are "fixed", ie. they don't change.

Apart from the fact that the instrument isn't calibrated and that I would like my 2700 to be significantly more stable in the slow rate, what worries me is that there is a significant change in the mV range between the values, which should be stable across the ranges. There is a ~6mV jump when switching from fast to med. What could possibly be the cause of this? As far as I can see from a Keithley 2000 schematic(the 2700 is almost identical to the 2000), there is no change in slope or anything when switching ranges. According to the manual, the only thing that changes is the number of power line cycles that the integration is preformed at. Since the slope is identical(I haven't looked at it with a scope yet), that just leaves some issue with controlling the charge/discharge of the ADC, or am I missing something here?

[Kleinstein]

First 4 quick questions:
What are the integration times and data rates for fast, med and slow rate ?
I assume that auto-zero is on for these readings - does the effect change between AZ and non AZ mode ?
Is the error more like a wrong scale factor or more like an offset ?
Is the noise within normal bounds ? For the 10 V test it could be noise of the source. A quick test would be with a short and maybe 9 V battery.

When changing the integration time there may be different cal constants involved. So it could be a cal problem.
AFAIK the K2000 may need a zero cal from time to time that can be done by the user.  However the difference is too large to be explained by a reasonable offset from the input (would be more in the low µV range).

The K2000 DC volt circuit uses first an AZ buffer, than a mux to change between functions and than a gain stage, just before the ADC. In AZ mode the mux switches between zero and the signal. If the gain stage for some reason does no settle fast enough, there could be a difference between the integration times.

Another possible problem with short integration times could be a rundown phase. For shorter integration the rundown part gets more important. If something is wrong there the error would show up larger for the fast mode and may be unnoticed for the slow mode. Such an error would also show up as a massive DNL error, especially in the fast mode. So a fast DNL test (e.g. watch a slow ramp/ decay like  20 mV to 5 mV in 10 V range over a time of a few 10 seconds - this could be produced from discharging a large cap thorough something like 10 M+10 K (measure at the 10 K)).

[petemate]

Hi Kleinstein, thanks for always supporting me!

The integration time for fast, med and slow is 0.1, 1 and 5 NPLC for the 2700. For the 2000 its 0.1, 1 and 10.
I assume that auto-zero is on per default, but I have not tried to change the setting. It seems to be only doable through remote programming.

I am not sure what to call the error. At the 10V range I see an offset of ~6mV between fast and med/slow, with the fast rate being the lower value. On the 100V range, I see an offset of ~60mV, but this time the fast rate is larger than the med/slow rate.

There is noise, no question. I don't have a super-stable voltage source that I could try to measure. Trying with a 9V battery, I get 9.16554(fast), 9.17490(med) and 9.17561(slow). This time the error-between-rates is slightly different, approx 9mV. Comparing with my Keithley 2000, I get 9.1824x(fast), 9.18244(med) and 9.18244(slow), which is what I would expect.

I have traced the input signal through the circuit path all the way to U163, which is the MUX you talk about, and it appears to be stable all the way to the input of the MUX. But then the MUX starts MUX'ing around and I no longer can verify the signal levels with any reasonable accuracy. The 2700 does not have a self-test function which allows the MUX to be put in a steady-state, so I can't see whats going on.

I haven't yet looked at the ADC rundown stuff. Will keep you updated!

Thanks again!

[Kleinstein]

If only settable via remote interface, I would consider the AZ mode default, at least for 1 PLC and up. The fast mode could be non AZ - so it could be more than just the speed changing. So it's at least worth checking the zero for the fast mode.

The gain versus offset question should show up for a zero reading and maybe a negative voltage. If there would be a problem with the zero reading, I think it would be noted. So it is more like a gain or maybe linearity error.

Being 60 mV off in the 100 V range is quite a bit. The 100 V range corresponds to the 1 V range with the 10 M divider up-front. Nothing should change there with the integration time. So an opposite sign error at roughly 10% of the range is not a good sign.

The 2700 should not be that old to expect old leaking caps, but it's usually a good idea to check  the supplies first, especially the analog part.

Is there really no self test with the K2700 ? I would consider this a big fail for a 6 digit meter, even if today a repair at component  level is hardly (at least not in the US, but could be still in China or India) economic if not done as a hobby. I would even consider a good power on self test a kind of requirement - just to make sure that easy (for the computer inside) to detect defects are shown. Even if this is only a DMM and not a jet-plane a wrong reading meter can pose some danger and it would be good to warn about a defect if possible.

[SilverSolder]

Just like with a jet plane - you need at least 3 DMMs so you can "vote" on the answer!

[bitseeker]

Is there really no self test with the K2700 ? I would consider this a big fail for a 6 digit meter, even if today a repair at component  level is hardly (at least not in the US, but could be still in China or India) economic if not done as a hobby. I would even consider a good power on self test a kind of requirement - just to make sure that easy (for the computer inside) to detect defects are shown.

From what I see in the docs, the power-on self-test checks the EPROMs (U156, U157) and RAM (U151, U152), then lights up the display segments and annunciators. Front panel tests enable you to check the keys and the display. If the front panel can't communicate with the main processor, you'll get a "No Comm Link" error. Other checks are done manually using the tables in the service manual.

It certainly doesn't seem to check much, unlike HPAK meters.

[petemate]

I agree that it sucks that there is no real self-test. I assume its because they couldn't fit it into the firmware. Its basically a 2000, but with some extra stuff added for e.g. time keeping and controlling of all those modules.

Anyway, I tried testing if there was some linear relationship between offset and voltage and there wasn't. And a 4 wire short thingie gave identical readings regardless of rate.

I then noticed something strange at around the 10V mark. Regardless of the rate, changing the input voltage did nothing. As in the display was being remarkably stable. Here is a result of the voltages read by the 2700 as a function of the voltage read by the 2000, when stepping the voltage around the 10V mark.


Note how there are several "flat areas", meaning that the 2000 did pick up a voltage change, while the 2700 did not. There should naturally be a straight line. I can't say if this has anything to do with the precious problem of the sample rates having different offsets, but its very strange behavior indeed. I don't know how to proceed from here. Guess I gotta break out the oscilloscope and start probing..

[SilverSolder]

That graph is reminiscent of A/D converter non-linearity...

[Kleinstein]

The curve looks like an linearity problem. With a K2000 I would suspect the TF245 resistor array, e.g. not providing the fine scale for the multi-slope rundown.

Does a similar effect also repeat at smaller voltages like 100 mV  (still in the 10 V range) - these could give lower noise.
It could than be interesting to compare the fast and medium modes over a little larger (e.g. 2-3 times) range: if it is a problem with ADC rundown, I would expect a kind of periodic structure with a shorter period and stronger effect for the fast mode.

[petemate]

Hi guys, sorry about the delay in replying, but I had to spend a long time doing some measurements.

I have produced these two huuuge graphs, which shows the linearity of the K2700.

This graph shows how the ratio between the K2000 and the K2700 change, as a function of input voltage:

Ideally, there should be a nice, horizontal line, perhaps with a an offsets when a range changes, but there is definitely a linearity issue as one nears the end of a range, as is evident at the switch from the 1V-to-10V range and from the 10V-to-100V range. Its also happening at the 100mV-to-1V range, but the noise floor of my power supply starts becoming an issue. I would assume that the effect is also evident at the 100V-to-1000V range change, but I don't have the equipment to test at these voltages.

Here is a log-plot of the above graph, which shows the issue better:

The linearity issue is affecting all sample rates equally, but I don't know if the issue is related to the issue that this post was originally about, which is about an offset between sample rates. However, the original issue is also evident from the offset between the rate plots on the graph: At the 1V range, the K2700/K2000 ratio is higer for the fast and medium rates, but at the 100V range the fast rate is significantly larger than the medium and slow rates.

(Note that I define fast as 0.1 PLC, medium as 1 PLC and slow as 10 PLC. The K2700 defines slow as 5 PLC, but I decided to keep the sample rate equal between the two units)



This graph shows a close-up of the last part of the 10V range and the beginning of the 100V range, with K2700 voltage plotted as a function of K2000 voltage:

The issue is exactly the same as in the previous post, but with the plateaus happening at slightly different voltages.

As you guys mention, it definitely looks like a sort of linearity problem, but the question is where the issue is. I'll start poking through the TF245 and see if there is anything weird. Kleinstein, you write that there should be a larger effect in the fast mode, if its an ADC run-down problem. Do you agree that it doesn't seem to be the case?

Anyway, it was quite challenging to do this setup. As I don't have a programmable PSU, I had to rig up a stepper motor to the rotary encoder on my PSU and have it rotate the know at a speed of ~1-5mV second. It took quite a long time to gather these samples, but it was a fun little challenge.

[Kleinstein]

For the linearity test there is no real need to check different ranges, especially the 100 and 1000 V range. These just use an Divider upfront.

For the graphs there are a lot of data points, but no good description what is shown.

My guess is it should be the difference between the K2000 to K2700 reading or maybe the ratio.

The error getting much larger near the end of the range is a first hint.
It's also remarkable that the errors seem to get larger for the smaller ranges - this would point to a problem not with the ADC itself, but more like something coupling back to the amplifier.

If it is possible to turn off auto zero, one could check the amplifier itself with the scope. So have a test signal (like a 5 V rectangular signal - maybe even the scopes cal signal) and follow the signal path.

[petemate]

The first two graphs shows the ratio of the voltage reading between the K2700 and the K2000, as a function of the K2000 reading. The last graph shows the K2700 reading as a function of K2000 reading.

It is possible to disable the auto-zero. I will try to do that and then trace the signal path.

[Kleinstein]

If I see it right the red curve is the fast mode. So in the last curve it even looks like the fast mode works a little better than the other 2.  the spacing between the  large errors (longer horizontal ranges)  is also rather large, some 100 mV. This does not fit an error in the rundown part. So the TF245 array part for the slow slope is likely not the cause.

A spacing of 200 mV and thus some 2 % of the range is even more than a full cycle of the run-up part. So it's a huge error. The curve also does not look like am amplifier issue - that would be more smooth, but not like repeated steps. This is more like something really bad with the integrator or the switching. The steps could be cases where the integrator reaches rather high voltages, maybe reaching saturation.
So if would be interesting to look at the integrator waveform.

[petemate]

Yes, the red curve is the fast rate. It does perform a bit better than the other two, yes, but maybe that can simply be due to the lower accuracy in the fast mode? The "plateaus" are there, they are just more rounded.

Anyway, I tried tracing a signal through the circuit:

I disabled auto-zero and applied a voltage that corresponds to a plateau. I could then confirm that the measured voltage doesn't update(or actually, updates ever-so slightly). I was able to trace the input voltage to both TP104 and TP105, while confirming that an 1mV change to the input voltage does indeed produce the correct 1mV change at both TP104 and TP105. I was also able to see this 1mV change at pin 10 of the TF245 network. But then it becomes a current and that makes it a lot harder to debug.

Now, when I say that it moves ever-so slightly, it means that when I change 1mV on the applied voltage, the signal changes by
one least significant digit, ie 10uV on the 10V range.


I had a look at the ADC, which does appear to be operating correctly. But the strange part is that the waveform is almost completely static. Usually, you'll se a variation due to whatever noise is in the input signal and in the instrument itself, but here there was none. I could find maybe one single change in the signal, but nothing that looks anywhere near an ADC sampling a real-life signal.

Here is a scope picture:


Ch1 is the ~9.8V plateau input signal, Ch2 is the output of U137, Ch3 is U153-1(discharge triggering NOR gate). It does appear as I would expect it to, but it just isn't moving around. Its simply "too stable". I can turn up or down the input signal and then the ADC waveform will again go back to its usual "moving" behavior", corresponding to moving out of the plateau.

I don't know where to go from here. From the look of it, the ADC is functioning, but the only thing between that and the correct input signal(at TP105) is the TF245 and some switching FETs, e.g. Q127 and Q131. I could try replacing those, but I have no idea why they would fail only at specific levels..

[Kleinstein]

A very stable pattern during the run-up phase is a first indication, that the problem is not related to the run-down.

A possible problem would be something like Q133 not conducting. If the voltage there gets too low, Q132 might turn on at the peaks.
The drain of Q133 could be a test point for this.
It only takes a little kick to keep the pattern stable.
There could be other effects like coupling through the supply that could stabilize some patterns.
This is a rather difficult type of failure - though in this case the error is really large - so the fault should be rather drastic. It's much more difficult to get really good linearity.

Another point to check: is the problem also happening at negative voltages ? This tests a different voltage range for the integrator output.

[petemate]

I checked Q133 drain and there is nothing going on. Its just a scaled-down version of the output of the output of U137, but pulled to ground whenever Q133 closes.

I then tried applying a negative input voltage:

As before, the ratio between K2700 and K2000:

Much nicer. No changes(a little bit around the -20V I'll ignore), except for some range jumps.

K2700 as a function of K2000:

So it would appear that there are only issues whenever there is a positive input voltage. Thats very odd. You talk about a the voltage range for the integrator?

[Kleinstein]

I am not that surprised that the error only happens for the positive side. It got better with smaller voltage (in a fixed range).

The output voltage of U137 is the output of the integrator. One thing that changes with large positive voltages is that the minimum at U137 out gets more and more negative.

Q133 seems to work - so no luck finding the fault there.

Another possibly point to check would be the supply for U137 and U138  so unusual spikes there (e.g. if the voltage at U137 out gets very low) could be a problem. The output of U138 could be interesting too, though a rather small signal. I would ideally expect a kind of rectangular (not very clean edges though) signal of a few mV amplitude.

[petemate]

Hi,

I had a look at U137 and U138:

Ch1 is TP105, Ch2 is U137-6 and Ch3 is U138-6. I think it appears to be normal.

I also had a look at the supply voltage to U137:

Ch1 is AC coupled and measured at U137-4.(with probe barrel connected to ground). Ch2 is U137-6.

There is some noise on the rail, but I don't see it dipping in any way and U137 doesn't go below approx -6V, so there should still be enough headroom. Note that the large noise at about +3.6ms corresponds to the communication of a value through the opto.

If you have any other suggestions, I'd very much appreciate it. Because right now I am basically at the point where I'd just start swapping out ICs.

[Kleinstein]

The output from U138 looks as expected. This suggests that U138 is working and the integrator is not somehow oscillating or ringing too much.

There is quite a bit of noise on the supply. I had expected less "noise" / ripple on the supply for a working meter, but this could be normal.
I have not expected large dips in the supply, the possible trouble would be more like larger spikes when the voltage at U137 out goes below some -5 V or so.

If I see the first measurement curve right, there could be a larger linearity problem in the 1 V and 0.1 V range. I am not so sure about this - it could be just noise of a problem with the source. If the linearity problem would really be larger in the lower ranges this would suggest that there is some coupling back to the signal before the amplifier, not just inside the ADC part. This would likely be something like a supply or ground issue.

Just to make sure the problem is not intermittent, is the problem still there ?  A failure (bad solder link) at Q138 would just fit very well to the observed errors.

I don't see any indication for a chip failure. Maybe Q132 could have issues (not blocking) - though this would likely cause more trouble.

Before replacing chips at good luck, I would more consider adding some extra decoupling (e.g. at U137 or U142 from +15 to -15 V) - in case it is a decoupling problem.

A slightly odd idea - more like a work around could be making the integration capacitor a little larger (e.g. add another 1-2 nF  low loss - like NP0 or PP) and this way reduce the voltage at the integrator.  This would increase the noise a little, but not much.
Even just as a test this could give a hint if the level from which on the trouble starts is more at the integrator or more at the amplifier before.

[petemate]

Hi, I thought it should be time for an update to my series on the Keithley 2700, since I finally got around to have a look at it during the Christmas holidays.

More than 6 months has passed since I last had a look at the unit, so I started by reading all the previous posts in the topic and verified operation. There is still a significant difference between fast and med/slow rate, as previously mentioned. There are also still these weird plateaus when one nears the end of a range.

By disabling auto-zero, I am able to read the signal all the way to TP105. At TP105 there are no plateaus, even though they still show up on the LCD display. So I assume that the plateau error isn't in any circuit prior to TP105. Some time ago I ordered a replacement AD711 off ebay as a test to see if replacing U137 would solve any issues. It didn't. I then looked at U138. Since the K2700 has a lot of OPA177's around(its constructed a little different than the K2000, with e.g. 2 OPA177's instead of a dual opamp(U139)), I tried switching around an OPA177 to see if the issue was with U138. It wasn't.

I then turned to the ADC itself(U165), since it would appear that the issue isn't anywhere prior to the ADC. I tried swapping the IC with the one in my K2000, to see if that gave any results. It didn't.

I then looked to the TF-245 resistor network, thinking that the current into the integrator could be the issue. Since its hard to verify the current in those circuits, I had initially skipped investigating this. I had previously tested that in my K2000(which had a broken TF-244 that I then modified), but thought that the swapping around could have caused some issues. I also noticed that the +/- 14V voltages were a bit low(+/-13.85 or so), So I removed the TF-245 and measured all the resistances, which appear to be within tolerance(or at least, no major outliers). Then I re-seated the IC.

Now comes the bad news: Investigating the resistor network bit more, I accidentally shorted pins 3 and 4 on U139-B(called U190 on K2700 and are non-inverting and -supply of an OPA177). And now the reference voltage reads about 0.6V. It would appear that the reference has been damaged, which sucks. Just another nail in the coffin of this project.

This of course lead me to look for another reference online. I can understand that the LM399 is "specially selected", but I don't know what that involves. Is it just a burn-in or is it some precision selection? I am thinking about buying a LM399H on ebay. As far as I can tell, the only difference between LM399H and LM399AH is the temp-co, which isn't much(1ppm/degC vs 2ppm/degC). Since this device will never be used for more than hobby use, I don't think that it makes a difference. Whats your opinion on this?

[Kleinstein]

The LM399AH is only slightly better than the LM399H.  The obvious difference is the TC specs, but there could be slightly more to this as the TC likely (for the TC with the heater of this is known) shows correlation with the absolute voltage. The abs. voltage can also correlate with noise levels. One can buy both the LM399 and LM399A with not that much price difference.

For the selection there are several parameters to look for: the TC, popcorn noise and drift during an initial burn in. As the 2700 is more like a lower cost meter, there may not be that much selection being used. The selected units probably have a slightly lower than typical noise and drift rate. I think the main part is to exclude those with noise or drift worse than typical.

Removing the TF245 array can be dangerous: these arrays are hard to get an quite expensive and they can be damaged or at least stresses by the soldering process. At least one would reset the drift to higher initial rates.

[petemate]

Thanks for your answer. How would I go about measuring the LM399(AH) for noise and drift? I do have a Keithley 2000, but it is uncalibrated, so I wonder if it even makes sense to use that to monitor the drift of a LM399..

[Kleinstein]

One could monitor the drift of the reference for the first times of use. So get a test point and measure with the K2000. That the meter is not calibrated does not matter, as long as there is no calibration (with adjustment) in between.  However the relevant time is mainly the time the heater is active - just in the turned off state not much will happen. So it may take a long time (e.g. 100 hours active) to see a change. It is than still not absolute for sure the drift is from the reference or the K2000.

The slight low reference voltage could be just a variation in the LM399. The absolute value has 2% uncertainty. So 1% off is not such a surprise.

For fist test to fix the linearity problem one could even use an LM329 (similar to LM399 but without the heater) - though higher TC could make some tests more difficult.

p.s. Added red "no"

[petemate]

That the meter is not calibrated does not matter, as long as there is calibration (with adjustment) in between. 

I am not sure I understand this.. If I measure with the K2000 for e.g. 100 hours, what would I need to calibrate and/or adjust?

[Kleinstein]

That the meter is not calibrated does not matter, as long as there is calibration (with adjustment) in between. 

I am not sure I understand this.. If I measure with the K2000 for e.g. 100 hours, what would I need to calibrate and/or adjust?

Sorry for confusing things. I somehow missed the word "NO". So it is Ok to have the not calibrated meter. It would be a problem if the meter gets adjust half way in between.