Absolute Divider Concept

[splin]

If there is no heat in the box then the thermals should be pretty good.

Well you'd hope so but standard switches aren't designed for low thermal EMFs so they could be absolutely horrible in this respect - you'd have to measure them to be sure.

SX relays are obsolete as at August 2011.

I mentioned them as another reference point for thermal EMFs for a relay described as low thermal. There would be no reason to use one when TXS2 relays are easily available.

The TXS relays are indeed quoted as you say, but for the non-latching relays. The coil heats the contacts and creates a temperature difference. An externally applied thermal gradient is a whole different (unspecified) state.

True, but I suspect you would be very hard pressed to create a temperature gradient, from an external source, across the contacts which exceeds that from the coil temperature rise of 5 to 10C specified in the datasheet. For a reasonable thermal design you should be looking at very much less than 300nV.

See this thread for some measurements - particulary chuckb's posts #30 and #37:

https://www.eevblog.com/forum/metrology/measurements-on-emf-error-of-switches/?all

Also see Andreas's post #38.

Chuckb's tests of an Electroswitch C4 rotary switch in post #39 is also interesting with thermal EMFs < 40nV. They start from £15 at Mouser.

[CDN_Torsten]

It's possible to minimize the thermal input of the non-latching relay coil by dropping the coil current to it's minimum hold current once the contacts have closed.  The hold current tends to be much less than the required current to start the contacts moving.

[chuckb]

This week I happened to be testing the leakage resistance of some C&K red 7201 series switches. They look like whats in your setup. C&K specifies over 1 Gohm insulation resistance and the switches meet that. With a Keithley 617 Electrometer I measured 400 Gohm from the common contact to case. From common to common of the DPDT switch I had 200 Gohm. An open switch contact measured 700,000 Gohms. You may want to evaluate if this will cause you any issues.

I have not noticed DIFFERENTIAL Temf over 20nV with the DPDT switches inside a metal box.


I also looked at some small NKK M-2022 DPDT switches and they were over 2,000 Gohms from common to case and between common terminals. The open contact was over 1,000,000 Gohms.


ITT Pamona Low Thermal EMF 3770 Binding posts have greater than 14,000,000 Gohms leakage resistance to the mounting plate at 100V excitation.

Hope this helps.

[Lesolee]

The elephant in the room is the resistors wired to the unused switch contacts. It was soooooo tempting to use the unused switch contacts as insulated standoffs.
PTFE standoff posts seem to be rare and expensive. However, it is far from ideal to couple the heat from the dividers into the switches!

100²/20K = 0.5W, which is a lot of heat to get rid of.

I would think the thermals would balance out due to the reversing action of the jig, or at least be visible as drift or offset in the readings. (There are 4 zero states, for example). More resolution needed to see the problem.

[quarks]

very interesting
unfortunately missed this completely, but now it is bookmarked
thanks for sharing

[Lesolee]

The TXS relays are indeed quoted as you say, but for the non-latching relays. The coil heats the contacts and creates a temperature difference. An externally applied thermal gradient is a whole different (unspecified) state.

True, but I suspect you would be very hard pressed to create a temperature gradient, from an external source, across the contacts which exceeds that from the coil temperature rise of 5 to 10C specified in the datasheet. For a reasonable thermal design you should be looking at very much less than 300nV.

I am not as confident in that idea as you are. The relay design could have a nicely balanced thermal path to dissimilar metals so that it gave a low thermal EMF to its own heat, whereas an external temperature gradient could just heat one brass/copper junction and give a big problem.

Just to get a feel for the thermals here I used my home-made micro-voltmeter (https://www.eevblog.com/forum/metrology/project-micro-voltmeter-design/45/, post 45 onwards) to measure the voltage across SW3 closed contacts. I didn’t use screened leads, but the result is pretty clean. Clearly there is not a HUGE thermal problem.



The µVM was wired up and on for 50 minutes before turning the PSU output up to 60 V. It is hard to say there was any actual thermal change. There is an ‘interesting’ jump at 9.5 minutes for no obvious reason. The power supply was turned back down to 0 V at around 12.3 minutes, and I suspect the dip is due to the common-mode change.

[e61_phil]

I used the Panasonic relays for my low thermal multiplexer. I tested a couple of cables (short).

I touched the cable for 30s with my fingers to see the thermal effects of the cable.

The voltage was measured with a Keithley 182. <300nV is achievable with the relays.

[Lesolee]

I used the Panasonic relays for my low thermal multiplexer. I tested a couple of cables (short).

I touched the cable for 30s with my fingers to see the thermal effects of the cable.

The voltage was measured with a Keithley 182. <300nV is achievable with the relays.

Can you spell out exactly what you are doing here please. It looks interesting, but I just can't follow the plot.
You touched the outside of an insulated cable?
Is the cable 4mm long?
What has "Fluke" got to do with anything?
You touched the cable for 30 seconds and the effect lasted for 10 minutes? 

[Lesolee]

This is a test of a 5 digit DVM module I got off ebay. It is the sister of the module used in the micro-voltmeter project. The discontinuity around 10 V is because the DVM auto-ranges at that point, so the half-value point suddenly gets an extra digit at that point. 60 digits translates to 6 digits which would have carried on the previous trend.



This other way at looking at the same data says the high/low ratio should be 2, but it deviates from that fraction by a certain number of ppm. If there was no linearity error then the curve would be a straight line along the horizontal axis. I created the red dotted line as a quadratic, where the x² multiplier was -0.0000313

It means we have an absolute INL reference.



[EDIT: This testing was done with a 100R pot in series with the linearity jig +IN connection, to fine tune the voltage to a nice value. This was an exceedingly bad idea! The source impedance creates a ratio error of up to 75 ppm. This probably explains some of the wiggles in the response (but not the overall shape).]

[Kleinstein]

Getting some error even near zero suggests that there is some offset error. The error is also really large.

That the error does continue even after the range switch, can have two reasons: The range switching can be more or less a software thing, not really changing the gain before the ADC. The likely used MCP3421 has some internal "Gain", using a faster input sampling for the SD ADC. The actual source of INL can be the divider before the ADC. In this case the main expected error would be of U³ type from a self heating effect and the linear TC of the divider.

[e61_phil]

I used the Panasonic relays for my low thermal multiplexer. I tested a couple of cables (short).

I touched the cable for 30s with my fingers to see the thermal effects of the cable.

The voltage was measured with a Keithley 182. <300nV is achievable with the relays.

Can you spell out exactly what you are doing here please. It looks interesting, but I just can't follow the plot.
You touched the outside of an insulated cable?
Is the cable 4mm long?
What has "Fluke" got to do with anything?
You touched the cable for 30 seconds and the effect lasted for 10 minutes? 

Sorry, I just wanted to show that these relays can achieve way less than 300nV.

The setup was the following:
The muliplexer has three channel. Two Inputs and a third one which is internally shorted. Which means one can switch Ch1, Ch2 or a short (Ch3) to the DMM.

I tested different low thermal cables (in the graph it was the Fluke 5440A-7002, which is one of the best). I used the "black" cable to short Ch1 and the "red" cable to short input 2. After some stabilization I touched one red and one black plug, to create a thermal gradient between the the plugs of the same color. This was done to simulate handling with the plugs. And yes it took many minutes to become stable again. But 30s ist quite long touching.

[Lesolee]

Sorry, I just wanted to show that these relays can achieve way less than 300nV.

The setup was the following:
The muliplexer has three channel. Two Inputs and a third one which is internally shorted. Which means one can switch Ch1, Ch2 or a short (Ch3) to the DMM.

I tested different low thermal cables (in the graph it was the Fluke 5440A-7002, which is one of the best). I used the "black" cable to short Ch1 and the "red" cable to short input 2. After some stabilization I touched one red and one black plug, to create a thermal gradient between the the plugs of the same color. This was done to simulate handling with the plugs. And yes it took many minutes to become stable again. But 30s ist quite long touching.

https://eu.flukecal.com/products/accessories/test-leads-probes-and-clips/5440a-7002
Ok, so the plugs are rated at 1.3µV/°C relative to tellurium copper. Assuming that is what they were plugged into, you heated each by 0.6°C and they took 10 minutes to settle down again. That seems fair. I thought you were saying you touched the middle of the cable. 

So what are the relays doing? Are you switching channels during the plot? Is it like 1 second per relay, then switch to the next? Are these latching or non-latching relays? Probably you have a good idea of what you are doing. I am old and slow, and it is difficult to figure out what exactly you are doing. (but it is clearly relevant and interesting)

[Lesolee]

The actual source of INL can be the divider before the ADC. In this case the main expected error would be of U³ type from a self heating effect and the linear TC of the divider.

If I apply linear voltage, 1 V, 2 V, 3 V  to an inverting active attenuator (so all the power goes into that resistor)
I get powers of 1, 4, and 9 power units
I then get temperature rises of 1, 4, and 9 thermal units due to the thermal resistance (°C/W)
This converts to shifts of 1, 4, and 9 ppm units due to the TC (ppm/°C).

Could you explain how you derive a cube law from this?

[Kleinstein]

The power at the resistors is proportional to the voltage square. This can shift the gain proportional to the voltage square. For the voltage error it is applied voltage times gain error this the Voltage as an additional factor. So a total of U³ * TC of the resistors * 1/R  / thermal resistance.

[Lesolee]

I should say a little about how this plot was measured. The voltage source was an ordinary bench power supply, ISO-TECH IPS603 (as it has been for all previous tests as well). This is quiet enough to get no least significant digit noise. The voltage output is not a pot, but a digital rotary thing, so it is nice to set, but doesn’t have quite enough resolution to get the last digit spot on. I therefore put a 100R pot in series with the power lead to the jig to tweak in the last digit.



I should also mention something about the presentation of these results. For a 58.00 V input I got readings of  +28.98 V and -28.98 V. The correct values are of course +29.00 V and -29.00 V. The positive reading is 2 digits low. The negative reading is arguably 2 digits high (more positive than it should be). I initially drew it that way, and it looks nice, with a linear sort of growth in the error. On reflection I changed it to make the negative error read as negative. In other words the magnitude of the reading is lower than it should be. I think this is more reasonable in the sense that it is the voltage magnitude which creates the power, which creates the temperature rise, which creates the TC error. This is symmetric. Making the curve look symmetric therefore seems more correct to me.

No doubt this will offend the technical sensibilities of at least half of the intended audience. I don’t claim that this is the “correct” way of presenting the data, I am merely stating what I have done. YMMV.

[KK6IL]

In the distant past I've used a insulated terminal which was a teflon bushing which a metal, probably brass, pin was inserted. Result looked about like the Keystone 11320, in stock at Mouser <https://www.mouser.com/ProductDetail/Keystone-Electronics/11320?qs=%2Fha2pyFadugenAX7QbWVFFz%252BtNpkTySy2A752Cfwr1QsAp2L0KjgLQ%3D%3D>
I really miss the days of printed catalogs where one could turn pages to find pictures of the item desired and see similiar items and prices.

[Lesolee]

In the distant past I've used a insulated terminal which was a teflon bushing which a metal, probably brass, pin was inserted. Result looked about like the Keystone 11320, in stock at Mouser <https://www.mouser.com/ProductDetail/Keystone-Electronics/11320?qs=%2Fha2pyFadugenAX7QbWVFFz%252BtNpkTySy2A752Cfwr1QsAp2L0KjgLQ%3D%3D>
I really miss the days of printed catalogs where one could turn pages to find pictures of the item desired and see similiar items and prices.

I looked for the same sort of thing and didn't find this one, so thanks.  (It is indeed PTFE bushing)
https://www.mouser.co.uk/ProductDetail/Keystone-Electronics/11320?qs=sGAEpiMZZMuHIszJOUpTOs%2FDGXBqE7ET

We used to use PTFE "cloverleafs", a tinned brass cup in a PTFE bush that you could insert wire leads into in order to solder them together without electrically interacting with the pcb, but still having mechanical support.

[Lesolee]

This week I happened to be testing the leakage resistance of some C&K red 7201 series switches. They look like whats in your setup. C&K specifies over 1 Gohm insulation resistance and the switches meet that. With a Keithley 617 Electrometer I measured 400 Gohm from the common contact to case. From common to common of the DPDT switch I had 200 Gohm. An open switch contact measured 700,000 Gohms.
...

I also looked at some small NKK M-2022 DPDT switches and they were over 2,000 Gohms from common to case and between common terminals. The open contact was over 1,000,000 Gohms.

The switches I am using here are from RS components (rswww.com). They are own-brand parts, RS part number 734-7022.

They quote insulation resistance at 1000 M min and dielectric strength at 1000 V.  I measured 1.2 T from a centre contact to the body using the micro-voltmeter from the micro-voltmeter thread (100M // 10nF NP0 across the input)

Open contact leakage 120 T.

Common to common is an interesting one. I was measuring around 3.8 T, but then I realised that this is a stupid measurement if the body is left floating. Grounding the body (naturally) killed the common to common leakage. I didn’t measure it properly, but at least a factor of 10 increase (ie > 40 T).

What I was most interested in measuring was the whole linearity jig. That is 4 switches, a pot, and 4 binding posts. Any connection to any point in the circuit connects to all points as far as leakage is concerned. The total leakage was around 800 G. That is adequate since my circuit diagram claims that 0.1 ppm linearity can be achieved with 200 G mismatched leakage from the mid point to the case.

[e61_phil]

So what are the relays doing? Are you switching channels during the plot? Is it like 1 second per relay, then switch to the next? Are these latching or non-latching relays?

Yes, exactly. The relays are switching between the three channels. One complete scan takes 12s. It switches the relays, wait 1s for settling and take 10 readings to average them from the Keithley 182.

The relays are latching ones and I routed both signals (HI and LO) through the relays to cancel out thermal EMFs.

The box is very ugly, but it works fine. (I had to use another red connector, because all the black ones from digikey were broken, but they send new ones immediately). On the pictures is another test with other very expensive Fluke cables. The "red" cable is used to short one of the inputs and the "black" is used to short the other one. Due to the wrong color on the connectors it is a bit harder to see what is really happening.

[Lesolee]

The box is very ugly, but it works fine.

Not ugly at all; mine look the same!



I forgot to upload the photo of the test. I applied the voltage to the case of the linearity box because the smaller terminals-and-electronics has less pickup area -- it is going to the current-sense node. It was not dangerous as I could only apply ±3V to the case since the micro-voltmeter is so sensitive. (I managed to get up to ±60 V doing the open-contact leakage on a single switch, since that leakage was up at 120 T).

[KK6IL]

PTFE double turret terminals available at Electronic Goldmine, 20 for US$ 5.00.
<https://www.goldmine-elec-products.com/prodinfo.asp?number=G24775&utm_source=Goldmine&utm_campaign=dcbac0a3e7-MAY23-2014_COPY_01&utm_medium=email&utm_term=0_15cb8e0368-dcbac0a3e7-60329037>

[Lesolee]

So we have a straightforward buy from Farnell
https://uk.farnell.com/itt-cannon/011-2049-040fb9/terminal-ptfe-test/dp/1347795
at £3.21 each (+20% VAT)

Then we have
https://www.goldmine-elec-products.com/prodinfo.asp?number=G24775
20 off for $5.00 (plus unknown international shipping cost)

The price difference is staggering, and may justify the international shipping. 

[Lesolee]

I used the Panasonic relays for my low thermal multiplexer. I tested a couple of cables (short).

I touched the cable for 30s with my fingers to see the thermal effects of the cable.

The voltage was measured with a Keithley 182. <300nV is achievable with the relays.

I just came across this Fluke short-circuit tool in an article "Watch out for those thermoelectric voltages!"



They want a short (brief) settling time so they are going for low thermal mass. I would suppose that this means it is more susceptible to drafts.

It is an interesting idea.

[e61_phil]

A couple of these shorts came with the 8508A and also with the 8588A.

[wiss]

You might like to take a look at this thread:

https://www.eevblog.com/forum/testgear/dmm-linearity/msg698554/#msg698554

Do read (or at least speed read) the paper linked in reply #3

The document of Jonas Wissting (Wiss) explains very well the mathematics that you need to arrive at better accuracy than each single resistor in the divider has. I think Andreas used a similar method with his ADCs.

So two of you think this WISS paper is good. Probably Affine Operators and Moore-Penrose pseudo inverse matrices are easy for you. Not for me. 

After reading through the whole paper I had no idea what he was doing. He did a test with 6 resistors in series. 6 resistors with unknown nominal values, unknown TCs, unknown tolerances. All we are told is that they were “lab-grade”. Is that a thing? Don’t cal labs usually use huge 4-terminal resistors in oil baths?

So having got to the end I had to go back and “reverse engineer” what it was he was actually doing. So he is measuring each resistor individually, then pairs, then triplets, and so on.

The claim is made that the resistors do not need to be of high absolute accuracy, which he has ‘proved’ by using resistors accurate to 0.5ppm relative to the group. There is no evidence or mention of checking for drift during any particular voltage setting, say the 1-resistor voltage set.

There is no mention of possible problems with the different source impedances as you move up the chain, or where the DVM guard shield is connected.

I suspect that it has been expressed in an overcomplicated way in order to impress the examiners. 

Nevertheless he has achieved some impressive results, using some impressive test equipment.  One can only hope that these results were not “cherry picked” out of less good sets of results.

Hi

I _only_ test "linearity" (in mathematics speak "affine") of the multimeter, if there is thermal drift, it will add to the non-linearity, as any other problem such as loading.
Data was cirtainly not cherry-picked, it was very tedious to read 7 digits of the multimeter, switch, wait for it to stabilize, repeat 
I found the resistors in a drawer at my old job, I do not have real data for them  (I could post a picture of the device)
You only have to understand what a pseudo inverse do, not how to reach it, there is a python implementation. It gives the least square solution to a problem with many solutions.
BTW, the 2 examiners were very confused, one did _not_ understand the mathematics (experimental physicist), the other (theoretical physicist, general relativity etc) did not understand the application  , it was fun