Scanner/Multiplexers for voltage references

[Andreas]

Hello,

more interesting would be wether the gain and reference drift could be calibrated out over temperature.

is there any possibility to measure the whole 0..5V range by using Gain = 1 and put the negative input to the 2.5V Vref?
So with a LTC1043 2:1 divider one could measure the whole 0..10V range.
Best would be a whole system calibration over temperature.

Edit: but I fear that the internal voltage reference is sensitive to humidity.

with best regards

Andreas

[dietert1]

My plan is to try and use this as a nanovoltmeter for scanning the voltage differences of a group of references in a round-robin manner. Want to compare results with the low thermal EMF relay scanner i made recently (see above). Still many steps to get there, like a good input protection for the ADS1263.

For measuring voltages up to +/- 14 V i will probably try and combine the second ADS1263 board i made with a programmable PWM reference, implementing a differential voltmeter. I want a precision DVM without resistor arrays and this may be a path.

Regards, Dieter

[Kleinstein]

I way to get low noise protection is using 2 back to back MOSFETs with a photovoltaic coupler to drive the gates and with a AC optocoupler for clamping and to turn off the input.
This is a bit like the input protection for the Keithley 2000 (ohms part) and 2182. It can be reltively low resistance.
For me 2x STU2N95 + TLP3906 + TLP290 work OK and limit the input current in case of a overvoltage to a reasonable level (< 100 µA).
There can be some thermal EMF (a few µV when heating a fet with the finger), but the FETs can be separated from the heat source (TLP3906).

To a limited extend (a bit high leakage) it can also be used for input switching. Compared to ready made photomos there may be lesser offset from thermal EMF.

For high speed pulses one may want some input inductors.

[dietert1]

OK, in order to try that scheme i ordered some parts from Digikey. TLP290 is "not recommended for new design", but Digikey have LITEON LTV-814 and LTV-844 that appear to be the same. STU2N95 is rare, so i ordered Infineon IPU80R4K5P7AKMA1.

Meanwhile i learned that noise and zero offsets of an ADS1263 depend a lot on PGA gain. At the highest gain G=32 i am getting 3 nV as stdev of the zero offsets. Noise when combining 2x 10 PLC measurements with polarity reversal is about 11 nVrms. For comparison a 3457A does a 100 PLC measurement including Autozero in 5 seconds and it gives 45 nVrms noise. At the same time the ADS1263 repeats the 2x 10 PLC measurement 12 times, so noise becomes 12 nVrms/sqrt(12) = 3.5 nVrms. Including the 10 channel scanner with calibrated offsets, there will be about 5 nV rms. Let's see how much input protection adds into this.

Regards, Dieter

[Kleinstein]

The FETs and optocouplers I noted are the parts that I use, they type are not special and alternatives are OK. The AC input optocoupler effects the current one sees in the fault case. For some reason the leakage specs for the STM FETs are relatively low - could be still just a point of the measurement. Just for protection and if not used as a switch the leakage is not critical anyway.

The noise level with the gain looks really good.  A relevant point may still be the input current.
Another possible source for trouble may be if the capacitance at the input effects the reading. This does happen with some AZ OPs and may also efect the ADC input. So one may want some filtering there (e.g. capacitance to ground and series inductance / resistance). Some part may be before the protection too. 

[dietert1]

LT MUX check with K2182A

The setup has one of the LT MUXes i made last year, with 16 low thermal shorts - soldered copper wire bridges inside scanner. Output is connected to a Keithley 2182A nanovoltmeter. A Nucleo32-L432 module with an ICL3232 is used to control the scanner and the nanovoltmeter by a USB connection to a W7 laptop. Besides the Nucleo32 has ambient sensors SHT35 and BMP388.
The nanovoltmeter collects 40 samples of 2 PLC with FAZ polarity reversal per sample in its trace buffer and sends the data to the host. This takes about 5 seconds per channel and about 80 seconds for one scanner round. Total number of samples is about 640 samples per scanner round plus some more data from the ambient sensors and the K2182A temperature sensor.
First diagram shows all samples from four rounds. Transients from relay operation are not visible.
Second plot shows 16 scanner rounds, this time with averages of each 40 consecutive samples. And there is an average curve from all 16 rounds. Once more the DIY scanner shows thermal EMF at a 1 or 2 nV level or less. This time there may be significant deviations.

What took weeks with a HP 3457A was finished after 30 minutes with the Keithley 2182A. HP 3457A five second measurements show about 45 nVrms noise, while a 2182A provides five second measurements with about 1 nVrms noise.

Regards, Dieter

[Kleinstein]

That is impressive low noise and also low variations in the thermal EMF for mechanical relays.

The overall offset of some 140 nV I assume is from the meter and maybe the cable there.

Not directly related to the scanner:
How fast is the FAZ in the K2182A running: is it 2 PLC with one polarity and then 2 PLC with the revese and than 2 PLC for the ADC zero ?

[dietert1]

Our second K2182A runs in a thermal chamber and with synthetic mains power. It exhibits noise of hourly averages of 0.18 nVrms, yet walks around like +/- 5 nV, so it needs continuous nulling. Maybe the LT Muxes i made are good enough to help improve below 5 nV. The first measurements looked promising.

When set for nominal 2 PLC, the K2182A executes two 2 PLC integrations for each sample if FAZ (polarity change) is on. So 40 x 2 x 40 msec = 3.2 sec total integration time. I don't think there is an ADC zero integration, as the conversion time totals 4.51 seconds including setup delays and delays from line sync. With three integrations per sample one would expect at least 4.8 seconds. I used the 2 PLC setting for lowest noise after looking at a diagram in one of the Keithley manuals.

Regards, Dieter

[Kleinstein]

Not using ADC AZ is a bit strange, as the non AZ mode has some low frequency noise, that can be visible even with a gain of 100 in front.
Not using AZ for the ADC may explain the extra slow walk around. A different range may give more hints on this.

With not AZ for the ADC part more gain for the front end could help.

In some cases the Keithley meters may use a bit strage type of AZ mode, like averaging over multiple readings. So the AZ part for the ADC may be faster than expected, though also not as efective. Still with just repeated 2 PLC readings (atually 2x2 PLC) there are not that many options.

[dietert1]

As far as i understand with polarity reversal there are two ADC readings ADC_P and ADC_M and the sample value 0.5 * (ADC_P-ADC_M) is independent of ADC offset. In the manual they mention about turning off autozero, but more warnings than explanations. The measurements above are with AZERO on. Maybe AZERO uses ADC_P+ADC_M to null the ADC.

Regards, Dieter

[Kleinstein]

If the description / crude plan is correct, the front end reversal only changes the sign of the offset. So one reading is actual signal + front end offset and the other is signal - front end offset. This should not correct the ADC offset. It is only the gain of the amplifier that reduces the effect of the ADC. There is still the amplifier before the ADC and the ADC itself with no zero correction. For the amplifer at the ADC input there may be a hidden AZ OP, I have not yet seen and than one may get away without an ADC AZ part. With high total gain the ADC drift would be less important - a really good amplifier may be sufficient.
There is still some residual LF noise from the 2nd amplifier and the ADC.

[dietert1]

Yes, that's correct. There is no polarity reversal, but an "exchange of parts" to simulate a symmetric input stage without offset.
Don't know why they didn't implement polarity reversal. Maybe the floating section coupling capacitance is large and takes current on polarity reversal. Then one would need a floating section that has its Gnd potential unaffected by input polarity reversal. Looking at the noise of 2SK170 below 1 KHz one would really like to have polarity reversal at a frequency of 50 or 100 Hz to supplement recalibration by the relay MUX.

Regards, Dieter

[Kleinstein]

The part exchange is just a slightly modified chopper amplifier. There is no significan analog low pass and the filtering is done digitally with the ADC. The method is kind of limited by the ADC speed and settling. The special switches (kind of Photomos) used for the input are also a bit slow.
Classical polarity reversal would need the switches to operate with the full input voltage and this add to the input current spikes.
There is a way to do a kind of polarity reversal for the ADC by moving around the low side (I use it with my DVM), but it really does not look like it is used here. This would allow a kind of AZ for the ADC withput loosing time, but it may not be compatible with the ripple from the front side AZ or at least add to the settling time.

Faster switching could help for the amplifer, but the ADC speed sets a practical limit to a 25 Hz cycle (2x1 PLC).

The strange part to me is more that the 2nd amplfier stage seems to not use an AZ op, but switching to zero with the MS ADC.
There are some AZ OPs in the circuit - maybe one is also correcting the 2nd amplifer, though it does not look like that.

[dietert1]

Meanwhile the LT Mux with K2182A revealed significant thermal voltages of up to +/- 6 nV on some channels. The offsets depend on the daytime and correlate with ambient temperature change. Our lab has a daily temperature cycle of about 2.5 °C, with rapid temperature rise after heaters turn on in the morning.
Due to construction of the LT MUX (see above) basic temperature gradients inside are near zero and cause thermal EMF of less than 1 nV on all channels.
As ambient temperature changes creep into the LT MUX and its relays, they cause additional temperature gradients proportional to speed of temperature change. I could measure the speed of temperature change (up to +/- 0.2 °C per hour) after mounting an additional temperature sensor inside the MUX and thus predict null offsets. Looks like the LT MUX+K2182A measurement can be calibrated to nV level. The diagram shows this for one channel of the MUX. All other channels exhibit the same behaviour with lower null offsets. In previous tests with a HP 3457A multimeter these effects did not show up, as that meter requires much longer integration times to reveal nV effects.
The internal temperature log demonstrates the quality of thermal isolation. The scanner alu box is inside an outer box with 3 cm of PUR foam.

Regards, Dieter

[dietert1]

Meanwhile there is a one week log with the setup. The image shows the LT Mux in its box, the Keithley 2182A nanovoltmeter and the Nucleo32 controller. Not shown is an HP 3456A used to log the MUX internal temperature, a USB to GPIB controller and a laptop computer.

There is a diagram to show all 15 calibrated channels over one week. The logs are stacked by 10 nV per channel, so channel 5 zero is at 50. The red lines are running averages of about one hour.

Calibration works in two steps:
1) Subtract measured null offset of channel 1, to continuously null the nanovoltmeter.
2) Subtract a linear term dependent on the speed of temperature change observed by the internal temperature sensor. There is a correlation diagram showing this relationship for channel 10 (same as in previous post).
First table shows the results of similar line fits for all channels. The fits yield positive and negative coefficients for different relays. Remaining offsets are all below 1 nV.
The residuals of calibrated measurements exhibit a standard deviation of about 2 nV (1.3 nV * sqrt(2) due to subtraction of channel 1)
Second table shows daily averages for all 15 calibrated channels. All 15 channels over all 8 days exhibit a standard deviation of 0.08 nV. As one scanner round takes about 80 seconds, there are 1000 rounds per day, so 2 nV / sqrt(1000) = 0.06 nV, so 0.08 nV is near perfect.

Next step is routing cables out of the LT MUX for all 15 channels and repeating test with external low thermal shorts.

Regards, Dieter

[dietert1]

As i did not have any free time to work on the setup i left it running. Now i have a one month log, with an overall standard deviation of each calibrated MUX channel = 0.3 nV. P2p is about 2 to 3 nV, roughly as expected.
I need to work on the determination of temperature drift after interrupts of operation. The math is incomplete during restart.

Second diagram shows a calibration check. There are two calibrations now, one from January (as above), one from February. They agree well. The per channel offsets are reproduced with a standard deviation of 0.09 nV.

Third diagram shows a Channel 1 log of the nanovoltmeter. That would be the result of the Keithley 2182A without the LT MUX. There is an uncertainty of 50 nV p2p and "events" that vanish after calibration by the LT Mux.

Regards, Dieter

[martinr33]

I did not see these papers posted before.

This one is on the Xdevs site:
https://xdevs.com/doc/_Metrology/qian2016.pdf
Interesting becasue:
 - it uses $4 relays (panasonic latching relays) and achieves very small offset voltages.
 - Noise appears comparable to the dataproof scanner system, below. +/-15nV
 - Uses terminal posts selected for similar thermal offset voltages
 - Pure mechanical connection to terminal posts.


NIST on a Dataproof scanner, showing about 30nV of noise per channel.
https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=32563

[dietert1]

The work i documented above includes a new method of calibration/compensation and it got explained in some detail. Another detail is a thermal clamp between nanovolt signal I/O and relays. In fact there are two thermal clamps, as shown above. Also i am using both relay contacts to implement symmetrical switching. My relay board isn't vertical but horizontal. My drive circuitry is USB without additional power supply. There have been some revisions to get there.
In my LT MUX design a delay of 5 seconds after switching a relay is not necessary.
In the next revision i want to implement a Kelvin connection for the internal temperature sensor.

Regards, Dieter

[Kleinstein]

The performance is really impressive. I had expected more trouble from using relatively normal relays for low level DC signals.
The thermal setup looks really logical: quite some separation the extra blocks.

With a good temperature sensor inside, it may be possible to also use some channels for thermocouples. So TC plugs connected with TC wire up to the contact block - these could be in parallel to the normal contacts.

[martinr33]

The general design approach for scanner control appears to be parallel controllers. However, this approach is wiring and power intensive.

I am thinking a serial shift register approach woud be better:
 - Fuve control lines, reset - power - ground - clock - bits
 - no operating current, just a quick pulse.
 - Very scalable, no meaningful limit to he number of relays
 - Board could be cuttable - design for 12 or 16 channels, and place cut lines so the board can be easily shortened.
 - there is a risk of multiple inputs being on simultaneously.

Two output channels might be good, ut they add another set of relay contacts.



Then, it needs a small controller. Arduino would be fine. The simplicity of the cable makes the separation very easy. Could also run off of a parallel port.

Another point, aluminum PCBs are now inexpensive. This approach would help with thermals. Surface mount relays would also be better.

[dietert1]

Yes, it's working well, yet for continuous logging i'd prefer a mosfet scanner.

With 5 seconds per channel and 5 msec relay coil pulse, the average power consumption is 2 * 5 msec / 5 sec * 4.5 V ² / 170 Ohm = 0.24 mW. One needs to operate two coils to step the scanner. Observed temperature rise of the internal temperature sensor above ambient temperature is about 1 °C when the unit operates inside its thermal box.
With photovoltaic mosfet drivers consumption would be about 3.3 V times 10 mA continuous = 33 mW. Roughly a factor 100 more than before. This would drive two contacts per channel like the relays (4 mosfets per channel).
For low leakage the couplers need to be close to the mosfets, but a temperature rise of 100 °C won't work. So the construction will be more complex, maybe with a peltier cooler to take away the "heat". Or give up on isolation and use mosfet drive circuitry similar to integrated muxes.

Regards, Dieter

[Kleinstein]

The single coil latching relays often get away with less power, but the driving with bipolar pulses is a bit more complicated. With half the coils it is still not that bad and there is a solution with N+1 CMOS drivers for N relais.

The PV drivers for a MOSFET gate can get away with a little less current. I get away with some 2 mA (2.2 K series resistor with 5 V supply). If needed as a pair, 2 diodes may be wired in series. The voltage at the PV driver is not the full 3.3 V, more like 1.2 V for an IR diode, though a few types may use 2 diodes in series internally. The resistor can be more with the control part.

The balance between heat and leakage is indeed a bit tricky.  Usually nV accuracy is not needed in combination with very high resistance and the MOSFETs are usually not super high isolation anyway. So the MOSFET based scanner would naturally have a bit more leakage, especially with many channels. For the heat of the PV drivers much would be about spreading it out / linking the different channels, as usually 1 channel is active at any time, just the position changes. A constant heat source is a smaller problem than one that is moving around.

In many cases one would not need that many channels, at least not the high precision ones. With slightly leaky switches more channels also come with a downside.  At least from my experiance one often has a combination of a few precision channels and than sometimes quite a few additional lower grade signals (e.g. temperature, supplies,...) to monitor. It can there make sense to have 2 (or even more) separate meters and scanners. The sometimes tricky point is combining the data.

[guenthert]

The general design approach for scanner control appears to be parallel controllers. However, this approach is wiring and power intensive.

I am thinking a serial shift register approach woud be better:
 - Fuve control lines, reset - power - ground - clock - bits
 - no operating current, just a quick pulse.
 - Very scalable, no meaningful limit to he number of relays
 - Board could be cuttable - design for 12 or 16 channels, and place cut lines so the board can be easily shortened.
 - there is a risk of multiple inputs being on simultaneously.

Two output channels might be good, ut they add another set of relay contacts.



Then, it needs a small controller. Arduino would be fine. The simplicity of the cable makes the separation very easy. Could also run off of a parallel port.

Another point, aluminum PCBs are now inexpensive. This approach would help with thermals. Surface mount relays would also be better.

     Well, the old (Keythley brown era) 705 uses (plenty of) shift registers.  With some luck you might get a good deal on the secondary market.  The low EMF relay cards however are less common and considerably more expensive.

[dietert1]

Five days ago i started another log with a low thermal relay multiplexer as shown above, this time using a different LT MUX unit. Unit three was finished very similar to unit two, ie. with internal low thermal shorts, with a shielded output cable and the special LEMO connector for the Keithley 2182A. It also got a 10 KOhm glass thermistor to track internal temperature.
The calibration obtained from this log resembles previous results, both for temperature drift and basic offset calibration. The standard deviation of hourly averages is 0.33 nV once more. If the calibration comes out similar between two units, that's another indication it will be stable.

Regards, Dieter

[miro123]

I have few questions.
1. Why the most people choose relays?
2. What are the relay advantages for this application ?  Thread description is - 'Scanner/Multiplexers for voltage references'
I asking this questions because I start building mux for my 6 channel lm399/adr1399 board- I choose different approach -> low power=low thermal EMF, solid state=small size. If this does not work I can ovalized it- I already made very good digital-TC oven. I have experience in digital control and it was easy task. With dual control loops and MISO control I achieve under 1mK stability in 24h
I have doubt that I'm in the right way - I'm curious to know why experience guys choose relay technology for such application

Thanks in advance