Prema 8017 (sick ?) vs. Tektronix DM5120 vs. HP 34401A

[sorenkir]

Hi,

I have tried to compare the stability of some of my bench DMMs:
HP34401A (6 1/2 digits),
Prema 8017(7 1/2 digits),
Tektronix DM5120 (6 1/2 digits) (see https://www.eevblog.com/forum/repair/tektronix-dm5120-6-12-digit-multimeter-(need-repair)/msg3076415/)

All of them are connected to the same (cheap) Plexiglas boxed AD584-M reference (AD584KH).
They are configured to perform measurements with the best resolution, with some filtering.
Log covers about 6 hours. Raw data plot (obviously, there is a slight calibration difference between them):



It seems that the reference stability is not bad. Temperature must have rise during the day.
The HP 34401A is very stable, the Tektronix is not bad, but the Prema does not behave as I would expect with "large" jumps at some points.
On this graph, the difference in ppm, centered around 0, vs. the HP that I considered the reference:



I don't know what could explain that, I have already changed the electrolytic filtering caps.
Any thoughts ?

Michel.

[The Soulman]

A bad connection from a connector, relay or switch? do you have the schematic?

[Kleinstein]

The jumps look a little like popcorn noise or a bad connector. Jumps of some 6 ppm look really bad, like a defect or error of some kind. Some smaller jumps, like 0.5 ppm would be normal for a LM399 or similar reference. So this could explain the curve for the TEK5120.

One could do a similar check with a short to see if the noise is more additive (e.g. amplifier) or more multiplicative (e.g. reference).

[Andreas]

Hmm,

are the jumps also there if the Premo is connected alone?
There could also be some charge injection from one of the other DMMs.

with best regards

Andreas

[maginnovision]

I'd check your 8017 analog board supply voltages and capacitors. That kind of movement is pretty excessive even at 2s integration. What was the integration time by the way, average or auto filter, 10/20/40?

[Dr. Frank]

Please pay attention carefully, what Andreas indicates!
On our MM 2020 we had exactly this effect. A HP 3458a DMM created current spikes by its AZ circuit, which influenced the other paralleled DMM, or the LTZ references. that may ressemble popcorn noise.
The 34401A is very similar to the 3458a, as it is a derivative.
Check your DMMs by only connecting one at a time!
Frank

[maginnovision]

Wouldn't you expect the effect to be much closer between DM5120 and 8017 if that were the issues? Especially the 12-13:12 excursion.

https://snapshot.raintank.io/dashboard/snapshot/kfxyRaEXbmXfhUVsGhqS4z2Ko4BO3927

Here is a comparison for you if the charge injection is not the culprit. It's a more stable source(ltz1000a) but should give you an idea of what should be. This is 2s integration plotted with DMM internal temp(The ambient temp swings are about double the range around 22C but due to an error when data was collected it's unsightly to view). This was also through a 3499B so there is going to be some amount of noise attributed to the multiple relays and connectors the signal moves through.

[maginnovision]

1ppm is 1uV/V or 1e-6. OP said he used an AD584 for ref so 1ppm is 2.5-10uV if I recall correctly. It depends which value he was looking at. My ppm values are based on the average of the first 100 readings as the nominal value which aren't included in my graphs. If you build your graphs after you can just find the average and use that. If the nominal value is zero I'd probably use absolute values in uV(depending on range). Alternatively if you've got a calibrated value for ref you could also use that.

[Kleinstein]

The ppm curve from the first thread used the difference/ratio (does not make much difference for values that change so little) relative to the lower blue curve. So the blue PPM curve corresponds to the middle curve. The curves look a little different because the raw data show limited resolution of the Tek meter and in the relative curve this is smeared out by the noise of the 34401 / reference.
One can recognize some of the features of the raw curve, though not so obvious to see.

Charge injection from the other meters may be a problem. It would indicate also some problem with the source / meters: either too much charge spike or too sensitive to spikes. Different source may also react different to current spikes - ideally they should suppress this quite a bit as spike is relatively high impedance.
The curve does not look much like a problem with charge spikes. In this case one would more expect some beat frequency from the different reading / chopper frequencies involved.

[sorenkir]

Hello,

Thanks to all for your inputs and suggestions!

To clarify, although I am not in expert in calculations, the first ppm graph was the difference between Prema or Tek value and HP 34401A value, considered as the reference value at any time, converted in ppm of the reference value and centered around 0.
I don't know if it is a valid method, but it highlights the jumps and quantifies them in ppm of the reference value.

Today, rainy day, ideal for experiments!

I have tried to measure 0V (short) for a few hours as Kleinstein suggested, and the plot did not show any abnormal jumps (calibration of the Prema is not good, for sure):



As it seemed encouraging, I have tried to connect the Prema alone to the reference. So far, only 3 hours of plot but it seems that Andreas's and Dr. Franks's intuition was correct: no more big jumps in the measured value.
Raw value, with the same scale as before :



As I don't know the actual reference value, which is slightly drifting in time, I have plotted the difference with the average value of the series, in ppm of this average value :



I am not sure whether the remaining 'noise' is acceptable, it might also be due to the reference, but it is far better than before!
I was a little worried by this DMMs (although it was a real 'score') because the analog board rev. A6 had been modified at some point in time to retrofit some of the later modifications found in the rev. A8 schematic.
And the modifications did not look very nice (new electrolytic caps, on the bottom left of the picture):



In seems in the end, it is not that bad.
The only issue is the calibration that is not good, and even screwed up in some ohm or amps ranges, but it is another topic!

Anyhow, I have learned something: I would not have imagined that one DMM could interfere in such a way with another.
Very interesting!

By the way, Dr Frank : what is the 'HP3458A AZ circuit' you are mentionning ? Auto-Zero ?

Michel.

[Kleinstein]

With a short the Prema meter still looks relatively noisy. I know it is not very lower noise like the high end 8 digit ones, but the noise level looks rather high some 0.5 ppm peak to peak would be 10 µV_pp or some 1.5 µV_RMS. That would be high even for relatively fast conversions like 0.4 second. The slow conversion should be lower noise - about a factor of 10 lower for 40 s.

The Prema meter uses an PLL for the clock. From other meters (Solartron, Datron) this is known to be a possible source of trouble. It is odd they use just the VCO from the 4046, while others use an extra LC based VCO and only the phase comparator from the 4046. This makes me wonder how they could achieve good performance at all.
PLL unlock could be a possible source for jumps.

[maginnovision]

I'd agree that is pretty noisy compared to mine. I get 1.01uV RMS noise with a short with 2s integration(minimum for 7.5 digits). The analog board doesn't look all that bad though(mine attached). I think the 5017's had a more finalized version of the analog board coming years later.

It's good to see that some of those larger excursions come from the 3458, I think that might have been difficult to rectify if it were internal.

[sorenkir]

Hi,

Noise values in µV depends on the range, right ?
Below, 5 hours plot of short on the input, 2s integration time, 3V range, AVG10, in µV :



RMS value is 0.52µV. It seems acceptable (to me).
Michel.

[Kleinstein]

With a short the nose depends on the range. In the 20 V range it should be mainly noise from the ADC. In the 2 V range it is likely mixed with noise from ADC and amplifier. The 200 mV range is likely dominated by amplifier noise.

0.5 µV rms for the 2 V range is quite a lot for 2 seconds integration.  The amplifier is some 2-3 µV_pp for 10 Hz BW. This would be some 0.1 µV rms for 2 seconds. The 0.5 µV_rms somewhat match the 20 µV pp seen in the 20 V range.  So this looks relatively noisy to me.

They claim 8 digit performance after 40/80 seconds this would suggest a noise level of not much higher than 1 µV_pp at 40 seconds and thus not much high than 4-5 µV_pp for 2 seconds. So it looks like the noise is too high be about a factor of 5.

[dietert1]

Recently i studied a little our Prema 5017. Analog frontend is revision A8.
Looking at the schematic i found those six 10 uF tantalum caps in the reference circuit:  C16 to C21. They are used with DG211 switches to derive reference voltages of +/- 3.5 V and 10 V in addition to the LM399 7 V reference voltage. At ambient temperatures of 25 °C those tantalum caps have leakage currents of about 1 uA. Times the 100 Ohm or so of a CMOS switch gives 100 uV, possibly a significant deviation.
The tantalum caps were replaced by Wima film caps that we had. The leakage of those film caps is about 100 pA, that is a factor 10 000 less. The measurement of our 14.1 V reference shifted up by 3.7 ppm and a slight day/night cycle observed before vanished, i.e. overall TC was reduced.

11 uVrms noise at 4 sec integration time is a lot. The Keithley has a similar noise measuring in its 100 V range.

Regards, Dieter

[dietert1]

About noise:
The Prema 5017 ADC implements a +/- 3 V range. For this range the front end passes the input voltage via resistors R24 to R33 (5x 22K = 110K).
For the range +/- 300 mV the front end uses amplfiers N8 & N14 with the voltage divider R109 & R110 for a gain of about 10.
These are the two electrometer ranges of the P5017. All other ranges use the resistor array R36 as input voltage divider. For the +/- 30V range that we use for measuring zener references the input voltage passes a divider 9 MOhm and 1 MOhm. The divider source impedance of 900 KOhm will be the dominant noise source. The 1000:1 tap of that divider gives a voltage range of 3 KV. The input protection circuitry puts a limit at about 1 KV.

I also checked the control signals for the DG211 switches in the reference. They are 50 Hz with a precise 2 usec timing margin (break-before-make). As the DG211 is specified with 1 pC typ. charge injection, this will cause a leakage current of 50 pA typ.

Regards, Dieter

[dietert1]

The Prema 5017 (and most likely 8017 is very similar) uses two secondary switchers in the power supply of its processor board, type LM2596 and LM2594. First one is the main supply with a toroid inductor. Second one is a helper for standby mode. It had an unshielded 1 mH inductor (open core, "bone"). I replaced that by a shielded inductor (closed core) and added 6 uH ferrite beads onto the three cable harnesses that run between the boards and the mains transformer. I also added 6x 100 nF film capacitors onto the mains rectifiers in order to suppress diode turn-off spikes.
This time calibration of the Prema 5017 shifted by 15 ppm.

Regards, Dieter

[branadic]

dieter1, did you manage to decrease the noise of 5017? It's not limited by the reference, but the ADC and/or its frontend.

-branadic-

[dietert1]

No, my mods were in vain. As i wrote above the ADC with its native range of +/- 3 V makes things difficult. When measuring a 10 V reference, the frontend uses a 9 MOhm / 1 MOhm divider. To make that work to 1 ppm is a piece of art, but one can't expect much better.
Would be interesting to use a calibrator and compare noise in the 3 V direct range to other meters, e.g. the HP 3457A. Maybe there is something to the noise shaping ADC.

Regards, Dieter

[branadic]

Why is it, that P6048, which uses the very same BK7, shows lower noise? Have I overlooked something?

-branadic-

[Kleinstein]

The 6048 uses the same control chip and a similar PLL, but the integrator, reference and input section is different. So the PLL section is likely not a big issue for the x017.

The ref. current is about 4 x larger  (5 K vs 22 K though with a higher effective ref voltage) for the 6048. In the 6048 the input section is current controlled and this makes low frequency noise of the integrator less relevant as the refrence is only connected part of the time. This especially helps in the noise test with a shorted input. Part of the ADC of the 6048 is proportional to the voltage and thus effectively adds to the reference noise. So not all the 6048 noise is obvious in the simple tests and it may looks better than it actually is.

Another possible issue that I see with the x017 meters is the switched capacitor part to create the -3.5 V ref. This looks like hard switching and this way sampling the relatively noisy LM399 reference not very often. This could add quite a bit of noise as with a short about 2/3 of the time the -3.5 reference is active. This may also create a linearity issue.
The TL064 OP-amp for the noise shaping part may also add some noise - though not sure how relevant that part actually is. With the relatively large integration capacitor the noise there can be an issue, especially for shorter integration.

If at all one would have to look at the 3 V range, not the ranges with an input divider.