LM399/ADR1399

[miro123]

Hello,
I've started an new software project.
it contains only four LM399.
The goals are
  - create model of each LM399
  - find best and worse of them
  - real time report of estimated voltages on each channel
  - much better performance than classical analog approach - averaging.

Project is more software related can be applied to multichannle LTZ1000. ADR1000s. I have limited knowledge in low level voltage measurements.- build two projects with ADS1256+PGA280 and just started with evaluating ADS125h02

Description is available @
https://github.com/mirtcho/LM399/wiki


Suggestions, remarks and critiques are welcome.

[dietert1]

Creating a Maxwell demon for voltage references is an ambitious project. You may need AI and a lot of example data for teaching it. Guess it starts by making one or more hardware setups to collect data. That is the state of my own work. I share your idea that similar methods will be applicable to different types of zeners. As Mr. Frank commented somewhere: "A good zener reference is constant but sometimes goes up or down by 0.1 ppm."
At the same time i know that all my setups contain errors, like connectors made of brass, imperfect Guard schemes and the like. The Arroyo temperature controllers i am using for some LTFLU references aren't perfect either. So it's important to keep up mood and resolve errors like the moving LM399 caps causing popcorn noise or RF from a WLAN router causing shifts.

Regards, Dieter

[Kleinstein]

With LM399 reference I quite often see jumps between 2 discrete level, like an upper and lower level, but relatively little noise while at the level. So if one could know which level is currently present.

For the measurements I would consider to only read the difference of each reference relative to the average of some 4 (maybe more) references.
This would lower the demand on the ADC(s) and one could monitor multiple channels.

[DavidKo]

I have checked the ADS1256 datasheet and I think that you cannot directly hook it to the reference (max input voltage should be+-5V) without divider.

[miro123]

I have checked the ADS1256 datasheet and I think that you cannot directly hook it to the reference (max input voltage should be+-5V) without divider.

The ground between two parts is shifted with 4,5... 5V. On such way the LM399 7v comes somewhere  in the middle of ADS1256 range.

[miro123]

With LM399 reference I quite often see jumps between 2 discrete level, like an upper and lower level, but relatively little noise while at the level. So if one could know which level is currently present.

Thanks for the input. This wil help me to build  LM399 model easier.

For the measurements I would consider to only read the difference of each reference relative to the average of some 4 (maybe more) references.
This would lower the demand on the ADC(s) and one could monitor multiple channels.

I don't understand this one. Why I need averaging? I would prefer to scale the setup from 4 to 8 independent channels instead of creating two -4*independent + 4 average. Average can be done in numeric way using the difference between signals.

[Kleinstein]

The ADS1256 is a pretty good ADC, but the S/N is still not much better than the LM399 and one would still need a reference for the ADC.  I have done measurements with a LM399 ref as the DUT and another LM399 ref at the ADC. In this combination alone one never known which ref. is at fault. It needs at least a 3rd one. The ADC noise still is part of the noise - even with a pretty low noise ADC.

So the logical point would be to subtract some common voltage (some 7 V) from all the inputs. There would be only a small rest left for the ADC to sample and look at. For the 7 V to substract one would need a prettly low noise ref (e.g. ADR1000 or similar).
The average votlage of multiple LM399 in the test would be a pretty low noise reference, though still not perfect. If you than see a jump in one of the curves, it chould be pretty clear who is at fault, as one curve would see a jump up, while the others would see a smaller (1/3) step down.

[bsdphk]

Look up "three cornered hat" in time/frequency measurement, you can do pretty much the same thing with voltages.

[miro123]

The ADS1256 is a pretty good ADC, but the S/N is still not much better than the LM399 and one would still need a reference for the ADC.  I have done measurements with a LM399 ref as the DUT and another LM399 ref at the ADC. In this combination alone one never known which ref. is at fault. It needs at least a 3rd one. The ADC noise still is part of the noise - even with a pretty low noise ADC.

So the logical point would be to subtract some common voltage (some 7 V) from all the inputs. There would be only a small rest left for the ADC to sample and look at. For the 7 V to substract one would need a prettly low noise ref (e.g. ADR1000 or similar).
The average votlage of multiple LM399 in the test would be a pretty low noise reference, though still not perfect. If you than see a jump in one of the curves, it chould be pretty clear who is at fault, as one curve would see a jump up, while the others would see a smaller (1/3) step down.

Sorry for my criptic description. My point is that I don't measure the absolute value. Absolute value can be measured onece per month or year since I dont own 8 1/2 voltmeter.
I measure only the differences between signals - worse case 7mv difference  - I don't see any problem to use any multichannel SD ADC. CMMR shoud not me an probleem too since Vref signals are almost DC.  The project will be software and mathematics oriented
https://github.com/mirtcho/LM399/wiki
Short oversimplified description for 3xLM399
1. LM399 Vz re marked ad V1, V1,V3
2. the number of diferences is N=(n-1)*n/2 in our case n=3 N=3*2/2
3.diferences are V1-V2, V1-V3 and V2-V3
4. with simple mathematics you can judge who lies from them.

You can make many this only with difference of two Vref's . Eg V1-V2 = contains the spectrum noise of both Vrefs.
Statistics can help me to distinguish good from bad LM399 samples
With only power cycling one device I can build model of power off/on hysteresis of one device.
Ideas are endless.
I'm aware that proper HW setup is essential.  I don't have experience with measuring low level signals.  I have many qustions about my HW setup.
Lets start with first one.
What are the challenges of connecting  +4,5V output from LM399 power domain to the ADS1256 GND? ->     7V- 4,5V ~2.5V half ADC range

     

[Kleinstein]

Mesuring only the differences makes sense.

Shifting the ground of the ADC relative to the LM399's is OK. It needs some care about the power on and power of phase, to make sure the ADC input voltage don't leave the allowed range. The current from the positive side is naturally limited as there are only some 1 mA to drive to reference part. With a much lower voltage this may go up to maybe 2 or 3 mA, but this should still be OK in the safe range.
The 4.5 V to shift the supply should still be low noise, so more than just a TL431.  The ouput interface would also need to use the same ground as the ADC.
So this would be more like having a separate negative supply for the LM399 low side.

A possible configuration could be using the LM399 as a negative ref. relative to some +10 V. So the ADC would be at ground and there would be a common voltage or some +9.5 V  for the common positve side of the references. The ~2.5 V to ground would also be used to provide the zener current with a resistor each. It would be just the question on how to generate the +9.5 V (likely something like 1.3x times the average or one of the reference).

[DavidKo]

What about using some low noise LDO as LT3094 for negative voltage? https://www.analog.com/en/products/lt3094.html (to be true, I'm not sure if it has lower noise in reality - it should have 0.8uV _RMS 10Hz-100kHz.

[miro123]

I've started breadboarding and I've ran to the following problems
positive:
 10V5 voltage is running noise free
 10V5 voltage seems to stable - Voltage vs temperature. short term and one week stability within the specs of resistors
problems: All of them are 7V output related
 7V - I buffered the LM399 7V with the second opamp of OPA2188 - result noisy 7V015 output.I cannot measure the drift noise is +-50..60uV
 7V - I added filter group - R6=0 Ohm - C1=1nF - result no change @7v015 output
 7V - I added filter group R5=1k C3=220nF/COG - result slightly beter but still unstable 7V015 output.
 7V - added output filter R7=100 C4=1nF - short term /10minutes/ stability is OK. Noise/unstability is gone too. But still the One day and one week stability is worse than 10V5 output

I perform measurement with Agilent 34970A

I have no idea what goes wrong. I have few questions
  - Do I connect properly the 7V015 voltage buffer using OPA2188?
  - OP2188 is comming from alliexpress - can be an defective component? - still strange why 10V524 works just fine with the first opamp
  - Is there any LM399 reference design for 7V?

I forgot to mention that I power cycle the seup every day. Start temperature inside is 20deg at the morning 30 degree ~17:00h

Best Regards
Miro


 

[Kleinstein]

The overall circuit looks OK. The filtering at the input is a bit unusual, but should not cause trouble.

C1 and R6 would not have much effect with only 1 nF for C1.  To have an effect it would more like take 1-10 Ohms and 1 µF as suggested for the LM1399.
The LM399 should be ok with the small capacitor.
The filtering with R5 and C3 makes absolute sense.  One could consider to take -INA also from there.
With only 1 nF C4 has little effect though it does no hurt.

A point I am missing in the ciruit is supply decoupling also for the OPA2188.

It is absolutely possible to get defect or rejet parts from Aliexpress, though fakes are more common.

[Andreas]

The overall circuit looks OK.

for me it looks very weird.

the Zener section is used as heater and the Heater is used as reverse polarized zener.
(usually pin 2+4 are the GND-Pins on LM399, and pin 1+2 belong to the zener)
with best regards

Andreas

[miro123]

for me it looks very weird.

the Zener section is used as heater and the Heater is used as reverse polarized zener.
(usually pin 2+4 are the GND-Pins on LM399, and pin 1+2 belong to the zener)
with best regards

Andreas

Thanks for the remarks, At the breadboard they are connected properly. I've just begin learning Altium.
BTW I still don't know why they called LM399's "Zener" pins but don't call the LTC6555 "bandgap pins"

[Andreas]

At the breadboard they are connected properly.

Strange. 50-60 uV I have only had on one device in my ageing box. Until I recognized that one of the heater pins was not within the socket. -> how much total current does your cirquit take? (Should be around 25 mA including Op-Amp at room temperature).

With best regards

Andreas

[RandallMcRee]

. . .
I have no idea what goes wrong. I have few questions
  - Do I connect properly the 7V015 voltage buffer using OPA2188?
  - OP2188 is comming from alliexpress - can be an defective component? - still strange why 10V524 works just fine with the first opamp
  - Is there any LM399 reference design for 7V?

. . .
Best Regards
Miro

Personally, I would never use an opamp bought from Aliexpress (or Banggood, or Amazon, or...). The usual opamp for your purposes would be either the OPA2189 or the ADA4522-2.

You can get the OPA2189 for a reasonable price directly from TI. (ti.com). Tests here on the forum show it to be a good performer. Otherwise use Mouser or DigiKey.

[miro123]

Personally, I would never use an opamp bought from Aliexpress (or Banggood, or Amazon, or...). The usual opamp for your purposes would be either the OPA2189 or the ADA4522-2.

I have changed the opamp with OPA2182 and problem disappears. That is what I have at home. I've bought 15pcs. of them from Mouser. All zero drift opamps are overkill for current breadboard. Isothermal zones are not optimal, no Hi-Z guards at all.

 

[r6502]

@Kleinstein

With LM399 reference I quite often see jumps between 2 discrete level, like an upper and lower level, but relatively little noise while at the level. So if one could know which level is currently present.

For the measurements I would consider to only read the difference of each reference relative to the average of some 4 (maybe more) references.
This would lower the demand on the ADC(s) and one could monitor multiple channels.

Do you have a reference for this - would be very interesting for me. What is the step height of the different voltage levels?

Guido

[maat]

Do you have a reference for this - would be very interesting for me. What is the step height of the different voltage levels?

Guido

I have attached a lovely sample of popcorn noise for you. The measurement was taking against a known good (LM399) reference using a KS 34470A. The output voltage of both modules is 10V. The popcorn noise in this case is about 0.5 ppm. Going through my logs of the last burn in run of about 50 LM399, I see about 0.25 ppm as the minimum and integer multiples of that number. The yield was typically about 60 % with no popcorn noise.

[Kleinstein]

The steps I saw are around 3-4 µV, so some 0.5 ppm of the 7 V reference voltage. Different referencey showed a different frequency of the steps. I have one that get on average 1 jump per minute and another one that is more like 1 jump per hour - so one can get away without the jumps for quite some time. Attached is a measurement curve measuring 2 references relative to each other. The jumps are likely all from one of the references.

[TiN]

Something showed up here to test.

[branadic]

"Pictures or it isn't true."

 

-branadic-

[miro123]

Creating a Maxwell demon for voltage references is an ambitious project. You may need AI and a lot of example data for teaching it. Guess it starts by making one or more hardware setups to collect data. That is the state of my own work. I share your idea that similar methods will be applicable to different types of zeners. As Mr. Frank commented somewhere: "A good zener reference is constant but sometimes goes up or down by 0.1 ppm."
At the same time i know that all my setups contain errors, like connectors made of brass, imperfect Guard schemes and the like. The Arroyo temperature controllers i am using for some LTFLU references aren't perfect either. So it's important to keep up mood and resolve errors like the moving LM399 caps causing popcorn noise or RF from a WLAN router causing shifts.

Regards, Dieter

I already start experiments with  6 channel prototype and I already start to struggle with small details :-)
1. From where do I get the PE -line? Do I need to make own PE?
2. Where to connect the Protective earth?- metal closet, insulation transformer
From where do I get me PE -line?
My whole setup is located in big metal closet = 2xHP3456+1xHP34960+1xPM2534+1X8842a + 6channel LM399/ADR1399 ready +ADS1256/STM32H743/ISO UART.
Power is delivered via 1KVA 230V/230V insulation transformer. Transformer itself has PE guard
3. Another question is where to put Raspberry PI? I use the Raspberry pi to collect processed data via isolated UART interface. The remote PC has acces via Ethernet port.

Enough questions for now.

As already mentioned the project is software based. The first step is to build reliable setup
Best regards,
Miro
 



 

[Kleinstein]

It is a bit unusual to use an extra isolation transformer. IFAIK the usual isolation transformers are build to power 1 consumer only. With more consumers connected it gets tricky with the protection and things can get worse than directly mains powered.

PE is mainly for protection against electrical shock and the earth found at the mains oulets or in rare cases with a separate connection. It depends if one actually has to connect the circuit to PE. In this case the references and ADCs would have no connection to the outside and may vrey well not need it, though it would not necessary hurt. There would be a transformer to power the references and ADC and STM32H... . Another (could also be the same, as the isolation is not absolutely needed, but it help with possibly EMI) transformer would likely power the Raspberry.
Ideally there would be a nearly closed metal case around the reference and ADC part with the case connected to the ciruit GND. This case could also be connected to PE, but it could also be isolated.