T.C. measurements on precision resistors

[Andreas]

Hello,

For those doubting the measurement results I recommend the reading of 2nd post on page 1 of the thread.
So battery powered devices, photocouplers and common mode noise filtering are all already done.
Otherwise I would have much noisier results as only the datasheet specs of the converter.
For the DUT I have done some improvements (thanks to Emanuels hint)
to keep the soldering junctions at same temperature (shown on later pages).

The absolute value of a resistor (or uncertainity) is not interesting me.
Neither for T.C. measurements nor drift observations.
I only need a refererence resistor that is stable during the measurement.
And why should the reference resistor (at constant regulated 27.5 deg C)
do any funny things even if its only a Z201? (Ok with drift this would be serveral days).

With best regards

Andreas

[Andreas]

Hello,

In the mean time I have done some measurements on 120 Ohms resistors.

Since my reference of the ADC cannot drive 2*120 Ohms = 21mA I have done a
voltage divider of the Z201#6 1K reference resistor against the DUT of 120 Ohms.
This also leads to a acceptable self heating of the resistors.

The drawback is that I have only around 0.5 V over the DUT.
And the noise level of 1uVpp from ADC over 1 minute integration time remains the same.
So ADC noise shows up as around 4 ppm (pp) related to resistance.

For the T.C. with LMS method many of these measurements are used.
And thus the noise cancels out mostly.

For the hysteresis, drift and T.C. with box method I am hard at the measurement limit.
So these values will be less reliable.

Attached the measurements of the first resistor

Z201 120R  #1 date code B0947-

19.04.2015: first measurement normal polarity
20.04.2015: 2nd measurement reversed polarity (check for thermal EMF)
21.04.2015: 3rd measurement normal polarity

LMS interpolation of 21.04.2015

A 0 =  5.47357940569535E-0001
A 1 =  7.98654305122316E-0001
A 2 = -1.46086280995968E-0002
A 3 = -4.29274706993186E-0004

So T.C. from LMS at 25 deg C is +0.80 ppm/K

The "box" T.C. is around 0.74 ppm/K including noise
and around 0.60 ppm/K from LMS interpolation

Relative large hysteresis on measurement visible.

With best regards

Andreas

[Andreas]

Hello,

further result Z201 120R #2 date code B0947-

23.04.2015: first measurement normal polarity
24.04.2015: 2nd measurement reversed polarity (check for thermal EMF)
25.04.2015: 3rd measurement normal polarity

LMS interpolation of 25.04.2015

A 0 = -7.64120247519562E-0001
A 1 =  6.22567646520884E-0001
A 2 = -5.21019214386767E-0003
A 3 = -2.94021440704217E-0004

So T.C. from LMS at 25 deg C is +0.62 ppm/K

The "box" T.C. is around 0.66 ppm/K including noise
and around 0.49 ppm/K from LMS interpolation

Obviously less hysteresis than on Z201_120 #1 on this resistor.

With best regards

Andreas

[HighVoltage]

Hello Andreas,

You are doing an amazing job in educating us on this subject.
As a relative new voltnut, I really enjoyed this thread and just wanted to say thank you.

[acbern]

I have had similar drifts on z-foild resistors. confirmed by Vishay AEs. For sure, a normal epoxy z-foild resistor cannot be used as a precision reference resistor. Too unstable; driven by moisture. Needs to be a hermetically sealed type.
Secondly, the question also is, what is the error budget of your setup per GUM. Have you ever done that analysis? Without, all measurements are meaningless.

[MisterDiodes]

Andreas,

Again, not trying to take away from your hard work, but here are some respectful observations why I do not agree with your numerical results.  All of the below comes from decades of experience in the lab:

Looking at your measurement device, I think you missed the point about optical coupling to get away from system noise - Yes you have an opto-coupler between the USB chip and the CPU, but that is not the requirement - You want anything that oscillates completely removed from the neighborhood of the test device altogether.  It looks like you have a CPU & USB chip on the same board as the ADC, with a noisy laptop and its internal switcher sitting very close to the resistor under test.  If that's what I'm seeing, that very easily will destroy measurements < ~5ppm if you're not very careful with your grounding setup.  I've never seen low-noise measurements at below 5ppm with such a setup with hobbyist boards.

What I'm taking about is that the DUT needs to be shielded and guarded completely. Without heavy shielding your test must be several feet away (at least) and isolated from any digital switch or oscillator of any kind (including un-isolated LTC2057 chopper amps) - and the only means of communication with test jig is over a fiber cable - usually in analog form telling us the resistance bridge is balanced.  It could be a digital signal on the fiber, but only if comm packets occur only when the device is not under active test.  We never use RF transmission in the area when testing DC components either.  Make sure all WiFi is turned off.  Its not even allowed in our building whatsoever as a security measure.

With a real standard reference resistor in its own shield / guarded metal box, then it is much more robust in the vicinity of digital switches, etc.

You do use a driven guard circuit, correct?  I don't see on one the schematic.  I have never been able to get good results on passive devices in the <5ppm uncertainty area without that.

Cheap quality test: If you are taking a resistance measurement, and you hold you hand close over the test jig - does the measurement change at all?  If so its probably not shielded & guarded enough. 

I know you are not interested about the absolute value of the reference resistor, but you are completely dependent on that reference resistor being -stable- during a test run, which it is probably not.  As DigilentMinds pointed out, that resistor value can be wobbling all over - and fairly quickly too - during your test.  Just the fact of your body and lungs exhaling (mostly water) coming within a few feet of that Vishay foil resistor can change the humidity enough to affect measurements. The PWW resistor won't care much, if at all.

I know you think that the Vref voltage isn't important during your test.  Be careful in assuming that.  The noise of your Vref still has a large effect with your measurement technique - and it has to be appropriate to the degree of uncertainty of your test jig.  Remember:  At the moment the ADC takes a measurement, it is using a roughly pure DC voltage on the sample & hold cap and comparing it to a shaky Vref.  The comparison is not instantaneous - and therefore any noise on Vref can affect any ratio measurement you are going after.  The noise may not be random in nature either - so trying to average ADC measurements in math can introduce weird effects on the results also.

Try this - If you think that Vref doesn't affect your ratio measurement, then add a sine or square wave to it and see what happens.  Say +- 1/2 volt (or whatever) @ 10Hz (sweep over a freq. range) and see what happens to your measurements.  Then add white noise to the Vref at a known level.  You can come up with a value in dB for the effect of both low frequency and higher freq wobbles on the Vref and how it affects the measurement outcome.

Other things to look at:
Study Linear Tech App note 86, and learn where you have to put in a solder compensation joint on your circuit.  You might have this already.

Then measure the effect of your soldered joints with a known, close-to-zero Ohms, gold wire (or gold-plated Tellurium) on your test jig and see what happens. Polarity reversals at least a few times per second.  You should be able to verify what effect your soldered joints have on the system, and then you'll learn why we use them as little as possible on the measurement path.  Depending on your solder technique, amount of solder, contaminates etc. you may also have introduced a small p-n junction diode into the test system.  When measuring under 5ppm, this is sometimes easy to have happen to you.  Don't forget the pathway from +Vref to your ADC and from ADC -Vref to your electron source.

Andreas: Back in early 2000's or so when the LTC2400 series came out, everybody got excited about doing resistance ratio measurements and comparisons like you are doing.  Nothing new.   For weigh scales etc. that don't need below say 10ppm accuracy, these ADC's are great and the ratio methods are sound especially in differential mode.  But trying to go after single-ended low noise measurements below 5ppm gets really tricky in a hurry.  Just because the datasheet says 24 bits in a large type face doesn't mean that your test jig is operating to that accuracy & certainty.  Just because the device says it self compensates for zero on the IC, doesn't mean you have zero compensation at the DUT - and that doesn't mean you are zero-compensating -during- the test at your DUT connections. 

Suggestion: Read up on how Wheatstone and Warshawsky bridges work.  Especially use the Warshawsky technique for resistances below 1k.  It will help you tremendously to learn how to calibrate your test jig and measure uncertainties at a low level and high accuracy.

For instance:  If you think you are measuring to 1ppm, then generally your system needs to do at least 10x better than that in meaningful resolution, even for ratio or TC measurements.  SO:  You want be able to take a resistance measurement to 1ppm.  On your test setup you purposely perturb the resistance under test by +0.1ppm, then by -0.1ppm, then nudge it by +- 0.2ppm, etc. In order to help validate your measurement technique, you need to accurately detect any sort of change at 0.1ppm steps, plus and minus.  Your reported results should be no lower than 10x the uncertainty of your test system capability; that's the rule that all labs I know of use.

And then test hundreds of devices that you have measured value and/or TC on other equipment.  Then I'll start to believe you when you say you saw 2ppm drift over a couple days. 

Even when you are measuring ratios against an unknown, uncalibrated "standard" resistor, this is how we do it.

Please consider learning how the bridge methods work, especially when measuring down into <10ppm area.  You will find out in a hurry why we don't use DVM's (and that especially means 3458's) without reference standard resistors attached in ratio mode.  Even then a quality resistance bridge will give you better results in the long run.

Bottom line: Bridges are really your best friend here for any serious discussion of resistor TC. 

[Andreas]

Hello Ken,

for the drift over days you convinced me: It could be the reference resistor.
Although I have had only around up to 3% rH changes during
measurement of the same resistor over a couple of days since using Z201#6.
So I really doubt the 25 ppm from humidity (which is for 0% to 100%) in my case.

So for Edwins 120 Ohms resistors I will use the 1K UP805#3
resistor as reference.
So I will have "no significant effects" from humidity.
At least a [possible] unstable Z201 will not spoil the
result of a UP805 resistor.

With best regards

Andreas

[HighVoltage]

Bottom line: Bridges are really your best friend here for any serious discussion of resistor TC.

Are there any affordable bridges available?
Can you name a few models and brands?
Thanks

[Vgkid]

https://www.hofstragroup.com/product/guildline-instruments-9975-direct-current-comparator-resistance-bridge/
Used :Guildline 9975(specs above,lots of knobs), esi242 different variants. Neither are very cheap
if you want new, ask guildline,measurement international...

[Andreas]

Hello
acbern, misterdiodes:

Of course a resistance bridge would give more accurate (absolute) readings.
The problem is: I will never buy one.

I would really appreciate if someone with a resistance bridge would do independant measurements.

But I fear that will never happen since it is very time consuming and
if you have a bridge I understand that you have to earn money with that
[or cannot disclose the secret details].

On my side I will continue with my ADCs and do my very best.
The target is to select among several resistors those that will fit best for my next LTZ1000 references.

I have made a evaluation on the measurements.
So typical spread is around 0.03 ppm/K for the T.C. at 25 deg from LMS interpolation for the 1K resistors with Z201#6 as reference.
With some exceptions for the drifting resistors up to 0.05 ppm/K. Over all standard deviation is around 0.013 ppm/K.

For the 120R resisitors the larger noise leads to about doubled uncertainities.
I have splitted the evaluation for UP805 and Z201 resistors.

The evaluation for Z201 120R resistors are still in progress.

With best regards

Andreas

[Andreas]

Hello,

further result Z201 120R #3 date code B0947-

30.04.2015: first measurement normal polarity
01.05.2015: 2nd measurement reversed polarity (check for thermal EMF)
02.05.2015: 3rd measurement normal polarity

LMS interpolation of 02.05.2015

A 0 =  1.48285225770232E+0000
A 1 =  2.15243292958098E-0001
A 2 = -1.19791700742452E-0002
A 3 = -3.89171660289224E-0005

So T.C. from LMS at 25 deg C is +0.22 ppm/K

The "box" T.C. is around 0.40 ppm/K including noise
and around 0.17 ppm/K from LMS interpolation

With best regards

Andreas

[acbern]

Andreas,

the Guildline 9975 is really a precise bridge, however it is probably in the 4k range. New bridges from Guildline or MI (I have got qotes) is in the 30 to 40k range. They beat the 9975 in accuracy, these are secondary standard bridges, so the price is justified. The 242 is not as precise as the 9975, but cheaper. I do not want to go into the details, but the 242 has its challanges (temp drift...). The 9975 uses a more stable principle of measurment. Precision is achieved in comparison, not so much absolute (the 9975 is intended as a 10:1 bridge anyways).

However, there are other ways to do precision measurements. You could use a Fluke 720A as a bridge device. The 720a manual gives information about that. A 720A may be available for arround 500 bucks. Also, you need a lead compensator and a Null-Meter. Another 500. And a voltage source, that only needs to have good short term stability, could be a good DIY source (LTZ1000 based or something similar). Best would be a 5440A, another 1k probably.

I am doing precision resistance measurements using a 3458A in voltage divider mode (e.g. a 10k reference versus a 1k DUT resistor, kind of a ladder calibration), with a very stable voltage source. This is based on the excellent linearity of the 3458A. I have done a GUM analysis on the error on both methods, and the accuracy is pretty good. It gets harder as the extremes are reached (between 10Ohm/1MOhm acc. is better 5ppm, 1k-100k arround 1ppm, using the 3458A). Again, you have to have a very stable voltage source though, with the right voltage, to use the full range of the 3458A.

If you want to do precision measurements, there is no way arround good gear. And a GUM analysis, so you know what tolerances you are looking at.

If you want, and as you are in Germany too, I could e.g. measure  a 10k resistor for you. The standards I have are calibrated to <0.5ppm acc. Send me a PM if you are interested. The resistor needs to be stable though (normal Z-foil with epoxy are not usable, but hermetic ones are pretty good (I have had drifts of only 0.2ppm pa). And temperature drift needs to be low (<1ppm/K; respectively drift analyzed).

[acbern]

Not sure I understand what you mean. I have two sets of wires, one connected to the front voltage inputs of the 3458A, one connected to the rear (or one could also just connect the cables to the posts of the two different resistors to be measured one after another). Need to wait for voltage stabilization. So the sense versus drive voltage difference is not of concern (if i understand you correctly). The 10V range is best for accuracy, but for small values, one needs to go down (and live with the then worse linearity), too keep power loss in the resistors low. Voltage reversal is achieved by changing the polartity of the source (the Datron 4808 supports that, it also gives the best results in terms of stability, compared e.g. to a 5440, which I actually like much).
The limit is at 1M, a 10M ohm resistor cannot be validated that way (I calculated about 200ppm error budget), as the impact of the input impedance is too high. The 720A is better there.

The guildline, according to its manual, has an accuracy of 0.7ppm from 0.01 to 0.1 ohms, which is impressive (and 0.2ppm up to 100k, 0.7ppm up to 1M), while the 242D is about 1ppm at 100ohms, and even 100ppm (!) at 0.01ohms (substitution mode, 242D manual dated June 71). It is 1ppm up to 10M though, where the 9975 is 7ppm. The 242D needs cal. In a nutshell, the 9975 is good for low values, the 242D is good for high ohms resistors, I would say. None of them is worth the mone for me, unless I find a chep one somewhere.
Have you, btw, done anything/experience with the Guildline 9930?

[acbern]

I am not doing it this way, I am not using the sense terminals, just the normal voltage inputs. I am making two independent measurements, voltage drops accross the resistor sense terminals, force terminals driven by the Datron, resistors in series. 12V on the higher resistor, 1.2V on the lower. Voltage reversal. This gives the most precise results, as it is using the linearity spec of the 3458A. Also the error calculation is a little more complex if you do it according to GUM, but with K=2 the resulting accuracy is pretty good. I am intentionally not using a low voltage switch, as it introduces other errors, and does not add any benefit.
The 34420A unfortunatelly, although it has good low voltage capabilities, has no transfer accuracy specification, the 3458A is unbeaten in that respect. I had contacted Agilent/Keysight for this, but, to make a long story short, no outcome. So the 34420A is not usable for this.

[acbern]

Well, the transfer accuracy spec really is what you need when you do relative measurements. And that is not specified unfortunatelly with the 34420A.

Exchanging front and rear may be a good thing to do, I need to figure that (impact) out. The 3458A, in its cal prcedure, does diffferents offset cals with front and rear, so part of the impact is probably covered. Question is, what inaccuracies remain.

It is true, for really high resistances, a high voltage and current measurement is the way to go, the 3458A will not help here. For low impedances, I tend to live with the limited accuracy (still less than 20ppm at 0.1 ohms). Cannot get the data with a 34420A, although it is probably ok for that purpose. Question is, what (GUM compliant) error budget do you assume.

[Dr. Frank]

Hello,

for 10:1 resistance ratios, for example to check my Hammon precision divider, (250kOhm:25kOhm), I also use the 3458A, similar to acbern.

I use a 5442A as an ultra stable 10V reference, and measure input and output voltages (10V and 1V) simply on the fixed 10V range, using the 3458As 0.02ppm linearity.

This 10:1 transfer should be better (about 0.2.. 0.3ppm) than the specified 0.05 ppm of value +0.05 ppm of range (= 0.55ppm total), as latter includes linearity AND other instabilities of the LTZ and reference divider circuitry, and is barn door wide specified.

Last weekend, I also experimented on the Ratio mode, in comparison.
It can also deliver superb results for DCV : DCV, but you have to be careful, as its behaviour is not fully explained in the manual.

First, the Sense (-) and Input (-)  must not differ greater than 0.25 V.
Second, The Sense reference input accepts +/-12V DC only, and always autoranges!
That means, it will change range from 12V to 1.2V , if  the 1V output of my divider would be applied to the Sense jacks, and the 10V to the Input jacks.

That's no big deal, if the instrument has been properly autocaled just before, otherwise you will have inconsistent results due to range change errors.

But you have no control over offsets on the Sense input!

It's better for the ratio mode, if Sense and Input use the same, fixed  10V DCV range.

Therefore, for Ratio, always apply the higher voltage, i.e. 10V to the Sense input, and the lower voltage, i.e. 1V, to the Input jacks, but chose manual 10V DC for the Input first!
In advance, you may measure the offset of the 10:1 divider output first , and let the instrument calculate the Null, and then enter ratio mode.

The manual does not cover this special setting, but  I strongly assume, that the instrument will really behave like this:

Sense = 10V DC range, Input = 10V DC range (manual, fixed)
Ratio Result = (Input - Offset) / Sense

As Sense is 10V, its offset plays no role here.

It is also recommended, to use statistics on the ratio function, to improve result and to control the stability.


In comparison, I did the following measurements.
The Hammon divider was not calibrated before the measurements:

1. Consecutive measurement on fixed 10V DC range, with averaging and statistics.:
Offset = 0.000 000 0V
U1 = 1.000 000 68V, sigma = 133nV, NSamp= 10 (one additional digit by statistics function!)
U2 = 10.000 011 2V, sigma = 155nV, NSamp = 10

Ratio = U1 / U2 = 0.099999956

2. Ratio function, Offset used on the divider output, ranges and voltages as above

Ratio mode reading of 3458A = 99.9999573E-03, sigma= 5E-09, N=16

So I conclude, that both methods may deliver identical results up to 8 digits (1e-8!!), but I feel much more comfortable, NOT using the Ratio function, as I can better overview, what I'm doing.


I would also avoid usage of alternative front / rear jacks for ultra precise ratio measurements, as this does introduce hard to control thermal voltages... The Jack-to-cable voltages are not included in the offset compensation/calibration !

And also, the front/rear switch is not very reliable any more, after these years..

Frank

[acbern]

Dr Frank,

I am trying to avoid the ratio mode using the 4 front terminals too, as the - to - terminal voltage is limited and also, as you state, i kind of feel like things are out of control a little.

Using front and rear voltage inputs is convenient, so i changed to that some time ago, and I would think that the offset issue is taken care of by the separate calibration of both inputs which the meter requires. One does have to wait for thermal settling, but actually I do not see, in my setup, any change after connecting the meters and waiting 5 minutes. The best anyway is to just use bare copper wire (maybe plated), thats also what Guildline recommends for their DCC bridges. While I am still using gold plated copper spades cable or the low emf pomona cables, i will do some tests with dual twisted pair teflon cable. The alternative, hooking up the wires to one resistor first and then the other also bares the risk of thermal voltages. so I have mixed feelings about this too. the thermalvoltages should anyway be eliminated by reverse voltages. I would alos recommend to go to the limits, i.e. 12V and 1.2V applied, helps accuracy. if the switch is worn out, that should show as a fluctuating, unstable voltage.

On a slightly different subject, I calibrated one of my 3458As recently, and it starts up with an error message: 'calibration: secure required'. obviously i had to unsecure the meter by entering the pass code wehn doing the cal. but how do i set secure again. the manual is not very helpfull, and somehow i do not manage to get that resolved. Not the frirst 3458A I calibrate, but somehow I fixed that last times (year ago or so), and do not recall how. must have just pressed some buttons intuitively, does not work this time, issue drives me nuts.

[Dr. Frank]

On a slightly different subject, I calibrated one of my 3458As recently, and it starts up with an error message: 'calibration: secure required'. obviously i had to unsecure the meter by entering the pass code wehn doing the cal. but how do i set secure again. the manual is not very helpfull, and somehow i do not manage to get that resolved. Not the frirst 3458A I calibrate, but somehow I fixed that last times (year ago or so), and do not recall how. must have just pressed some buttons intuitively, does not work this time, issue drives me nuts.

Dear acbern,

things you're doing not so often, can drive you crazy!

It's described in the calibration manual, and as I don't want to do a real calibration right now, I'm no 100% sure, that I'm telling it correctly, now.
(I'm also well over 50, therefore I remember the last CAL only dimly.)

Well, if I figure it out correctly, there is no explicit UNLOCK / LOCK process needed.

You simply type in CAL 10, 3458, and the 10V calibration will be done with assumed 10.000000V reference.
Without the default password '3458', this calibration will fail.

As said, an explicit LOCK command is not necessary afterwards, or can nowhere be applied.

SECURE 3458,1234 would only change the password to 1234, and would activate the password safety for the CAL command.
SECURE 3458,0 would simply deactivate the password for the CAL command, so only the default calibration value (10V or 10kOhm) would be needed.
SECURE XXX,YYYY, ON /OFF would also secure /unsecure the ACAL feature.

This simple mechanism is also, what I remember very dimly from the last CAL I did, after you kindly measured my VHP202Z resistor..

It's different to the LOCK switch at the rear of the 5440B, isn't it?

PS: Probably, this confusion is caused by the HP34401, or other SCPI instruments.
This instrument has an explicit UNSECURE and SECURE command, which you have to use before and after the 35 points (?) calibration.
As the 3458A only has 2 calibration points (nearly), such an elaborate process makes no sense, I think.

Frank

[acbern]

thanks, and what you say is what I thought too. I did the cal as you describe, pretty straightforward. Command, voltage level, password (3458). I calibrated 0, voltage (10V), resistor (10k), and  ac (scal), the usual stuff (except maybe scal, but was necessary). all went ok, and meter works fine, but when I power up the meter, it always comes up with 'cal error, secure required'. so there must be something else...

[ManateeMafia]

What if you rewrite the security code? ie. SECURE 3458,3458

I remember getting this message after changing the NVRAMs. Once I did a SECURE 0,0 the message went away.

Syntax SECURE old_code, new_code [,acal_secure]

[acbern]

You made my day! I entered:

SECURE 0,0
SECURE 0,3458

Issue gone. Maybe just typing the second line (SECURE 0,3458) would have done it too.
That should now have set the cal security code back to 3458 after having removed the initial non-locked state.
Pretty logic overall, and many thanks

[Andreas]

Next resistor Z201 120R #4 date code B0947-

this one came directly out of my long term ageing experiment (load life) with 100mW intermittend power at room temperature

03.05.2015: first measurement normal polarity
04.05.2015: 2nd measurement reversed polarity (check for thermal EMF)
05.05.2015: 3rd measurement normal polarity

LMS interpolation of 05.05.2015

A 0 =  4.60338549126336E+0000
A 1 =  6.92576655745229E-0001
A 2 = -8.92617577082719E-0003
A 3 = -7.33433859833421E-0005

So T.C. from LMS at 25 deg C is +0.69 ppm/K

The "box" T.C. is around 0.71 ppm/K including noise
and around 0.59 ppm/K from LMS interpolation

Together with the measurement of 06.05. total drift at 25 deg is around 6ppm through 4 days.
In the last days it was rainy so humidity increased from 50% rH to 58% rH.
I think I really am in urgent need of a humidity insensitive reference resistor
so that I can tell wether it is the DUT or the reference which is drifting.

With best regards

Andreas

[Kleinstein]

The high values of apparent drift tend to show that there is a problem with the measurement setup. Also the curves with reversed polarity are visibly different. So I think some more tests and improvements of the measurement setup are needed. The noise also looks quite large to me - I would expect less noise in a well shielded setup.

At least there is no more need for even more curves of questionable quality.

Humidity likely is one more parameter to influence the resistors, and possibly also the measurement setup, through leakage currents. So I might be helpful to read the humidity data as well.

From the pictures at the beginning of the thread, the rather long unshielded and dangling wires don't look very good. They may cause EMI problems, by operating as an antenna. Also they make the effect of reversing the polarity questionable, as moving the cables may also change thermal gradients. As a thermal cycle is rather slow, it might be better to do the polarity reversal more often within a cycle (e.g. every 30 seconds - to really separate possible thermal EMF problems.

[Andreas]

Hello,

Also the curves with reversed polarity are visibly different. So I think some more tests and improvements of the measurement setup are needed. The noise also looks quite large to me - I would expect less noise in a well shielded setup.

I do not find a 0.03ppm/K difference (for the 1K resistors) is a dramatically difference with reversed polarity.

Noise is at the datasheet spec of the used converter. With 1uVpp (9uVpp / sqrt(85 samples)) at ADC input at 500mV voltage over 120 Ohms gives around 4ppm pp noise when regarding that each measurement consists of 4 single voltages.
By the way the noise is still better than many 6.5 digit instruments

There are always 2 possibilites: shielding or (common mode) filtering.
I use filtering + galvanic isolation.
EMI effects are checked to be below noise level of 1uV.

So I might be helpful to read the humidity data as well.

From the pictures at the beginning of the thread, the rather long unshielded and dangling wires don't look very good. They may cause EMI problems, by operating as an antenna. Also they make the effect of reversing the polarity questionable, as moving the cables may also change thermal gradients. As a thermal cycle is rather slow, it might be better to do the polarity reversal more often within a cycle (e.g. every 30 seconds - to really separate possible thermal EMF problems.

Thats what I already did in the overview sheets last week.
The problem with humidity is that it needs several days of time constant until it gets fully visible.
I blame the drifts (> 1ppm @ 25deg for 1K resistors, or > 2-4ppm for the 120R resistors) on humidity.
I will see if I can get better with the PWW-resistor as reference resistor.
Otherwise I will have to use a hermetically resistor.

I also plan a polarity reversal every measurement cycle.  (< 1 second)
(just to check EMF: I do not expect changes on T.C.)
But this will take some time.

With best regards

Andreas

[Andreas]

Last resistor Z201 120R #5 date code B0947-

also came from ageing experiment but was switched off with Z201_120#4 on evening of 02.05.

07.05.2015: first measurement normal polarity
08.05.2015: 2nd measurement reversed polarity (check for thermal EMF)
09.05.2015: 3rd measurement normal polarity

LMS interpolation of 09.05.2015

A 0 =  1.47787395857863E+0000
A 1 =  6.39202276575331E-0001
A 2 = -9.37087943635444E-0003
A 3 = -9.72385941996082E-0005

So T.C. from LMS at 25 deg C is +0.64 ppm/K

The "box" T.C. is around 0.65 ppm/K including noise
and around 0.53 ppm/K from LMS interpolation

No significant drift (1ppm) during the 3 days

With best regards

Andreas