Scanner/Multiplexers for voltage references

[Kleinstein]

Latching relays are a quite attractive solution:
They only need relatively little power for a short time (e.g. 100 mW for 10 ms and thus on average not much power).
The isolation is very good, especially the typical values.
Isolation between control and signal.
Low capacitance and thus little EMI coupling.

Semiconductor switching is desirable in same areas, but these have more leakage, a limited voltage rating, ESD sensitivity (may lead to leakage going up over time) and still need some power.
The main factor is how fast to switch and what votlage.

For the temperature it is not about the absolute temperature, but about avoiding temperature gradients inside the relays or semiconductor switches to cause thermal EMF.
So for a scanner circuit temperature regulation does not really help. It is more about a careful thermal design to reduce the gradients from heat flow through the cables and from the control part.

With a limited voltage and common supply / ground, like in a multi channel reference it absolutely makes sense to use semiconductor switching. the LM399 and ADC1399 are limited accuracy anyway and it is not about the 10s of nV.

[MegaVolt]

Semiconductors have a problem: they don't work without electricity. You need to ensure that the sensitive outputs are not wired to ground or between each other if power is lost.

[dietert1]

Also the idea is that a relay implements switching between metal parts, while mosfets involve semiconductors, where the Seebeck effect is known to be much bigger.
Anyway, my study (like others before) indicates that one can reduce temperature gradients to such low levels that the residual thermo voltages are in the nV and can be predicted and subtracted.
A mosfet solution would be preferable for durability, but for the time being i made the relay scanners. Unit 1 has been continuously scanning a setup with three voltage references since last June, with a total of about 310 000 scanner rounds, where the relay spec in datasheet is up to 50 000 000 operations (Axicom V23079-B1201-B301). So it should last about 78 years. Of course with 16 relays the first one wil fail earlier, maybe after 5 or 10 years.

Regards, Dieter

[dietert1]

Recently i used a Keithley 2700 with a 7706 multiplexer for resistor array TC measurements.
Maybe it can also be useful for scanning voltage references (difference mode). After a little experimentation i got this setup, for the time being using the 2700 front panel inputs: The meter setup is DCV range 0.1 V, "Slow" with autorange and filter off. Every 5 seconds it takes a trace of 38 samples. That takes about 3.5 seconds. The remaining time is used to transfer the data to the RS232 host and calculate the median value. So there is a result every 5 seconds. Standard deviation over all 5122 results in about 7 hours is 67.4 nV. Need to repeat this with ambient temperature logging and the 7706 as scanner.

Regards, Dieter

[dietert1]

Meanwhile i wired the 7706 scanner with 9 low thermal shorts and ran a 24 h test. One scanner round takes 45 seconds. Channel 5 gets used for continuous offset calibration. In order to reduce noise from the calibration channel i am using a running average of 5 scanner rounds, i.e. 225 seconds.
The diagram includes some data taking failures caused by a bad setup of the USB-RS232 adapter (CH340/341).
The table shows that with continuous recalibration each of the other 8 channels yields a resolution of about 50 nV. P2P appears a little more than expected due to the data taking breaks. The scanner needs some minutes to settle.
The residual offset of each channel from the calibration channel is between -88 nV and +20 nV, with an average of -21 nV over all 8 channels. These offsets are stable to +/- 5 nV when looking at partial data.
I think the Keithley 2700 with a 7706 plugin is a resonable solution for scanning voltage references in difference mode.

Regards, Dieter

[bobAk]

What relays are there?  I have a board 7701 2 pcs.

[dietert1]

The relays used in the 7706 have a red cover and are labeled "NEC Japan EC2-4.5SNJ". There is a number 389930, maybe a date code. They are single coil latch type relays, mouser have datasheet of Kemet version:
https://br.mouser.com/datasheet/2/212/1/KEM_R7002_EC2_EE2-1104574.pdf.
I mounted the multiplexer in the lower bay with the upper bay empty and with its cover on.

Regards, Dieter

[bobAk]

So the relays are the same.  Thanks. I will need to take measurements with a 7701 connector and a homemade cable.  Connector contact material is phosphor bronze.

[dietert1]

Meanwhile i solved the USB crashes and there is a complete 24h log now. Results are very similar to what i got before.
The day-night temperature cycle of about 2 °C causes offset variation of about 125 nV.
The third diagram shows the 8 input channel logs with continuous recalibration and stacked by 100 nV steps in order to show how well this is working. Total vertical scale is 1 uV and applies to channel 1. These curves involve a 5x running average.
With 5 second measurement time one gets about 47 nV RMS. This is 0.005 ppm of 10 V - good enough for monitoring a setup of multiple zener based references.
In this test the K2700 performs very similar to a HP 3457A + 44492A scanner plugin that tested with 45 nV RMS previously.

Regards, Dieter

[aronake]

I plan to use an Agilent 34970a with a 34904a matrix switch card to have 4 3458a continuously monitor 7 voltage references where all meters measure all voltage references. This will also be a way to monitor the 3458a.

The 34970a and the 34904a are not metrology grade equipment. The 34904a is specified to have a thermal offset below 3 uV. If actual offset is close to 3 uV with much variations on the relays this would not bee good enough for my intended purpose. But usually specifications are very relaxed compared to real performance. Obviously I wanted to test this.

Test setup:
- Well warmed up 3458a, 34970a, 34904a.
- Around 1.5 meter Canare L-2B2AT cable connected to the 4 rows and 7 columns (I did not test the last column). Shield of all cables soldered to ground on the matrix switch card. Canare L-2B2AT is a tradeoff between being thin, being shielded, being flexible, but unfortunately not PTFE but PE insulation. But PE is not too far from PTFE when it comes to electrical characteristics.
- Reason for the length of the cables is that this is the length i will need once i put all in "production" so get a test of the whole system. Not too much care on keeping the cables free from electrical interference. But this will be similar to "production" environment. I do not have enough space to make this great.
- All row wires connected to 1 3458a, and from there connected to one more 3458a. The second 3458a used to validate the first.
- All column cables shorted.

The test:
- Python script to:
- Switch through all switches of the 34904a.
- For each switching of 34904a relay do 7 measurements at NPLC 100 with both 3458a. First one 3458a measure, then the second, then the first again.
- ACAL on 3458a for every third round of switching of the 34904a. The first round after each ACAL discarded in final analysis, as the 3458a take some time to settle in after ACAL.
- Total 20 runs across all switches were made.

Results:
TEMF averaged 123 nV. Max 249 nV and min 32nV.

I took average nV measurement of all measurements on all switches and mapped out as how they are located on the card.

See nV switchmap attached at end of this post.


Generally the result came out much better than expected. 249 nV will not really be much seen on 3458a at 10V. 249nV is around 1/10th of 34904a thermal offset specification. Also cables and electronic interference included here. It is also very clear that different relays have different amount of temperature gradient. The transformer of the 34970a is located close to where the highest thermal EMF can be seen.

I also wanted to check settling time of TEMF after a relay has switched. To do this, I grouped all measurements in which order it was made since relay switched and took average of each group.

Measure   nV
1           117
2           118
3           119
4           115
5           119
6           117
7           116
8           117

Conclusion here is that no settling in time can be observed. TEMF is same for first measurement since relay switched to last.

All in conclusions:
Positively good result and the 34970a / 34904a are good enough for my needs when it comes to having many 3458a monitoring each other and many voltage references.

Improvements that could be done:
Best nV measurement device i have is 3458a. It showed to be good enough to measure what I wanted to measure, so the reasonable approach would be to be satisfied with this. The more fun approach would be to get a nanovoltmeter (2182A or 34420a). That will probably be where I will go next.

Ideas in improvement on 34970a/34904a:
It is clear that the relays close to the transformer show quite a bit more TEMF. I will do some test putting copper sheets on top of the relays to have heat spread out more evenly and see if that reduce TEMF.
 

[alm]

- For each switching of 34904a relay do 7 measurements at NPLC 100 with both 3458a. First one 3458a measure, then the second, then the first again.

Nice test and write-up, thanks for posting! If you're not aware, a trick to do long measurements more efficiently is to set the 3458A to single trigger mode, then trigger each 3458A (using TRIG SGL command) so they start the measurement, and then poll each meter until the ready for instruction status bit is set. See this code for an example of reading both non-blocking (max_time = -1), polling for a certain amount of time (max_time > 0) or a plain blocking read (max_time = 0). The advantage of polling is that it keeps the GPIB bus free for other thing. This is especially a big deal with resistance readings with OCOMP and DELAY which can take minutes.

- ACAL on 3458a for every third round of switching of the 34904a. The first round after each ACAL discarded in final analysis, as the 3458a take some time to settle in after ACAL.

Interesting, I have observed this for resistance but not for DCV. But then I normally use the 3458A on the 10V range, so maybe it's only the lowest DCV ranges?

[aronake]

Hi Alm,

Thanks for comments.

- For each switching of 34904a relay do 7 measurements at NPLC 100 with both 3458a. First one 3458a measure, then the second, then the first again.

Nice test and write-up, thanks for posting! If you're not aware, a trick to do long measurements more efficiently is to set the 3458A to single trigger mode, then trigger each 3458A (using TRIG SGL command) so they start the measurement, and then poll each meter until the ready for instruction status bit is set. See this code for an example of reading both non-blocking (max_time = -1), polling for a certain amount of time (max_time > 0) or a plain blocking read (max_time = 0). The advantage of polling is that it keeps the GPIB bus free for other thing. This is especially a big deal with resistance readings with OCOMP and DELAY which can take minutes.

Background to why I have one 3458a to wait while the other is measuring is that I first had them to measure at the same time. But then saw some mysterious result that I thought could have to do with the meters somehow causing some interference with each other. So then changed. I later figured out this likely was related to making measurements straight after an Autocal.

My usual method for nonblocking reads looks like this:

while not (MySTB & (1<<4)):
            MySTB = 3458a.read_stb()
            time.sleep(2)

What you point at looks better. Do you have some example code where this is used?

- ACAL on 3458a for every third round of switching of the 34904a. The first round after each ACAL discarded in final analysis, as the 3458a take some time to settle in after ACAL.

Interesting, I have observed this for resistance but not for DCV. But then I normally use the 3458A on the 10V range, so maybe it's only the lowest DCV ranges?

As mentioned in my writeup I did one ACAL for each 3d round of scan across all switches. In the graph below (not very nicely made) I have marked when ACAL happens (3 times for this part of the measurement). Each horizontal line is 100 uV and I added some offset to actual measurements to keep measurements from the two meters more apart. The faint" digital looking line" is what relay I was meaning on.

One meter seems to take a jump upward and the other a bit lower after ACAL compared to for the relay sweeps for where no ACAL was made. But it looks like it is very small, maybe 20 nV on the one that jumps upward and maybe 10 nV for the one that jumps downward. 20nV when measuring 10 V would be 2 counts on the last digit at 8.5 digits so hidden well below the noise floor. So for 10V measurements ACAL stabilization should not matter. And 3458a not really the right tool for nV level measurement. But what to do if not having anything better at hand (yet)?

This behavior looks very similar for all other times ACAL (maybe 7 more) was done over these measurements.

Purpose of this test was to see TEMF on the scanner, so do not have too good data on ACAL stabilization time, so not very conclusive.. But thanks for giving me an idea on something to dig into. I have 5 3458a, so not a huge sample from a statistical point of view, but still quite some 3458a to test. Its intresting that the two I used here seems to move in different direction after ACAL.

[alm]

What you point at looks better. Do you have some example code where this is used?

Here you have an example of starting a measurement on two 3458As and then polling both meters (in series) until they're ready.

Purpose of this test was to see TEMF on the scanner, so do not have too good data on ACAL stabilization time, so not very conclusive.. But thanks for giving me an idea on something to dig into. I have 5 3458a, so not a huge sample from a statistical point of view, but still quite some 3458a to test. Its intresting that the two I used here seems to move in different direction after ACAL.

I misspoke, I did not observe any deviation after ACAL, but I did observe a deviation after running the TEMP? command, which I run regularly to check if the internal temperature changed enough to require ACAL. When measuring 10 MOhm standard resistor, I observed about 3 ppm drops in the reading the sample after executing TEMP? (see attached graph, blue is the resistance reading and red is temperature reported by TEMP?). I did not observe any difference in the second sample, so I just discard the one sample after TEMP?. I haven't observed any effect on the sample after ACAL that was above the noise floor, but I indeed haven't used the 3458A down to these levels.

[aronake]

Interesting. Thanks for this.

I realized i misspoke too. 20 nV which was kind of my rough estimate of ACAL TEMF drift of the most impacted meter is 0.2 of last digit at 10V and 8.5 digits. So not only below noise floor but also below number of digits that can be read. So seems fair to say that DCV measurements can be done straight after an ACAL with no impact. Apparently not resistance though from your findings. I will investigate this further.

[aronake]

Time for some modifications to see if the 34904a can be improved. Wrapped the relays in two layers of copper foil, added some insulation and added grounds copper foil as shield around the card. Previously had the 34904a and a 34901a multiplexer in the 34970a. Have added one more card now to fill out the space.

A bit too many changes to know what change will make a difference, but running tests now.

[aronake]

The results are now ready. After doing the mentioned modifications: copper wrapped relays, some insulation, add one more scan card to fill out the machine, reduction of TEMF turned out to be 35 nV on average. What is more important that the improvement was bigger on the relays that previously showed more TEMF bringing down the min to max from 217 nV to 182 nV.

Results table below with per relay TEMF in nV, after mod, before mod and difference.

The setup was already good enough for my purposes before, but better is of course better. It is not certain which of the 3 modifications made a difference, but I could think that all helped.

[branadic]

I already raised my question in DIY Low Thermal EMF Switch/Scanner for Comparisons of Voltage and Res. Standard, but did someone with such a low thermal emf scanner already implemented a remote control for a fully automated setup? I was told the following already:

nV scanner: it is powered via the SMA connector (center positive). When you apply 5VDC, you can use the Reset and Next buttons to select the channel. If it is connected to the DVM's 'voltmeter ready' connector, it advances automatically, but be sure to set some slow measurement time - 1s or more.  Otherwise it won't have time to charge it's internal capacitors and won't have enough energy to switch the relays. Best performance is obtained when You delay the moment of taking measurement by 5...10s from the activation of the relay coil.

My meter doesn't have such a 'voltmeter ready' connector, hence why I'm asking for remote control.

-branadic-

[dietert1]

The low thermal EMF scanners i built have a nucleo32 STM32 controller (USB connection to host). The STM32 firmware also implements ambient sensors and a transparent RS232 link for the Keithley 2182A nanovoltmeter. In some sense the controller represents a meter with a scanner. These things were easy to do.
Meanwhile i learned how to implement GPIB using STM32 controllers. This improves measurement time (sampling rate) in relation to data transfer time. What is "missing" is the precision temperature readout of the scanner internal temperature sensor (glas ntc) that was previously implemented using a HP 3456A meter and got voltage deviations to below 1 nV. I made some prototypes including enclosures, but then i burnt one of the ADS1256A ADC modules and i continued with other stuff.

Regards, Dieter

[branadic]

Just in case someone missed the thread, here are very first results of a measurement on the DIY Low Thermal EMF Switch/Scanner

-branadic-

[dietert1]

Using the same technology as i recently showed in the "Experiments with reference ovens" thread, i instrumented one of the low thermal EMF scanners from 2021. Here i have some results from a first 24 hour test.
Temperature is measured with a 10 KOhm thermistor and a 10 KOhm metal foil reference resistor both inside the MUX. I am using a ADS1256 module with an ADG736 CMOS multiplexer for bridge voltage reversal. The half bridge is wired
with separate force and sense (5 wires in total). In order to fit this into the existing DSub-9 connector the relay driver was supplemented into a SPI like serial interface using a 74HCT595. The STM32L433 device includes a GPIB controller for the Keithley 2182A nanovoltmeter and a RS232 interface in addition to the USB CDC link. It also monitors ambient using I2C sensors.
The controller calculates temperature drift every 5 seconds and outputs a running average with a time constant of 100 seconds. In the diagram i marked the extrema of the LTMUX temperature curve (red), in order to check that the calculated drift result crosses zero at the same time. Notably the ambient temperature crosses the LTMUX temperature curve at the same time, except at the beginning, when the MUX started with its own temperature after working on it. The MUX is inside a thermal insulation box that causes a temperature time lag of about 2 hours.
The temperature drift determination is needed to numerically compensate residual EMF of the relays (about +/- 5 nV).

Regards, Dieter

[dietert1]

Meanwhile i started using the USB device with the scanner. The scanner currently has three input cables on channels 9, 10 and 11. I checked the wiring with an external voltage source, yet for the purpose of calibration they have a short on them. All other channels have shorts inside the scanner.
Once more i used channel 1 for continuous recalibration of the K2182A offset. With three days of data results are similar to what i got in 2022 and the calibrated scanner with the Keithley 2182A operates with < 1 nV stdev when averaging 10 scanner rounds (800 seconds), on all 15 channels including the three channels with external short.
Next level of evolution might be substituting the USB host connection by a fiber ethernet interface, powered from the K2182A.

Regards, Dieter.

[dietert1]

That ethernet fiber interface is on its way, but will need some time.
For the time being i ordered a no name USB isolator at ebay. It has a ADUM3160 inside and is capable of full speed, which is good enough. It also includes a small isolated DC-DC "BSPSU B0505S-1W". The isolator runs the STM32 based GPIB and scanner controller without any additional 5V supply.
Today i took a closer look. Capacitance between both isolated ends is 42 pF, similar to a small two chamber mains transformer. After unsoldering the converter with only the ADUM3160 left there is a coupling of 3 pF - similar to a shielded mains transformer. While running without load, the DC-DC injects about 200 mV at 50 KHz with lots of RF content into the isolated side and the 1:10 scope probe. Output voltage was measured inside the GPIB controller to be 4.7 V. Enough to work but not ideal.

Regards, Dieter

[Kleinstein]

Those cheap DCDC converter blocks are often rather poor with common mode injected signal. It may be worth to just disable the DCDC and have your own 5 V supply from the meter side.

For controling your own scanner part, one should not need a high data rate and no need to go to GPIB in between.  USB - UART (RS232) and than isolation for the UART should be enough.

[dietert1]

The GPIB port is for collecting the nanovoltmeter measurement data, if that was your question. The USB connection works as host connection for data storage. And yes, i am currently studying a small, low common noise 5 V supply. The Kniel C5.3 i recently got is good, yet a bit large. Hopefully the ADUM3160 behaves without isolated side supply.
Maybe i will also test some USB fiber link but probably it will become an ethernet fiber connection. The fiber transceiver and the DP83822 PHY consume about 1 W. I plan to reuse that design for the Prema 6048 remake.

Regards, Dieter

[dietert1]

Today i asked myself what may be the effect of the common mode signal emitted by the USB isolator resp. the built-in isolated DC-DC.
The scanner, scanner controller and nanovoltmeter form an island with one protective earth connection in the nanovoltmeter mains cable used for the meter and the controller metal enclosures, the connection between them being a 0.5 m GPIB cable. I measured the voltage across this cable, from one enclosure to the other one. I used a shorted out probe next to the USB isolator, picking up the stray field of the DC-DC converter as a trigger signal (yellow trace). As the measurement signal was about the same size as random noise, i used averaging to suppress random noise. In the beginning the GPIB cable had a 6 uH ferrite on it that i removed later. With ferrite ringing along the GPIB cable was 20 mVpp. As expected it reduced without ferrite, to about 7 mVpp.
The scope power supply was ground lifted for this measurement and the scope was grounded via the probe from the nanovoltmeter.
Then i recorded the signal associated with GPIB readout and audible scanner stepping every 5 seconds, this time without averaging. Random noise was about 24 mVpp. Then the GPIB operations (about 200 msec duration for transfer of 40 readings) stick out as four blocks with about 50 mVss and the large spikes of about 100 mVss are likely caused by the scanner operations.
My conclusion would be that the USB isolator is good enough and that one better avoids a ferrite on the GPIB cable. Don't know yet how to determine ringing on the nanovoltmeter input signal.

Meanwhile i got enough calibration data to determine the residual thermal EMF on my three external shorts to be about 1 nV/K of ambient temperature. The shorts are made by soldering the ends of the copper signal wires in the shielded cables to each other.

Regards, Dieter