Noise cancellation circuits for voltage references

[KT88]

@ maat: My comments and circuit were only about a (at least in LTSpice) working lowpass with leakage mitigation.
I agree on your statement about the reference source itself - it wont give more accuracy unfortunately.
If I'm not completely wrong, an increased precision should be possible (assuming the filter works as simulated).
Precision would mean in this case that a better short term stability would be achieved. This could help with some kinds of measurement like noise analysis or transfer.

[macaba]

With the 1/f corner of LTZ1000 being around 0.3Hz, there is still a practical engineering challenge to be had with achieving a DC stable (vs. temperature) 0.3Hz LPF?

[maat]

If I'm not completely wrong, an increased precision should be possible (assuming the filter works as simulated).
Precision would mean in this case that a better short term stability would be achieved. This could help with some kinds of measurement like noise analysis or transfer.

Indeed, that is quite true, although the latter could also be achieved using digital prost-processing (assuming the transfer standard does not have a similar 1/f component). It is fairly straight forward. Noise analysis is a little more tricky due to the convolution involved and the assumption, that the DUT is most likely not showing white noise.

With the 1/f corner of LTZ1000 being around 0.3Hz, there is still a practical engineering challenge to be had with achieving a DC stable (vs. temperature) 0.3Hz LPF?

Do not forget, that the 1/f corner frequency moves up proportionally to the white noise component. The formula is f_c = h_{-1}/h_0. The h_α is the power law coefficient as defined in [1]. h_0 is the coefficient of white noise. h_{-1} the coefficient of flicker noise.
So if you push down the white noise spectral density (in V^2/Hz) by a factor of 10, the corner frequency moves upwards by a factor of 10. The LTZ1000 is spec'ed at something around 50 nV/sqrt(Hz), so lets assume a filter should get this down to 1.5 nV/sqrt(Hz), this would mean that f_c moves to 0.3 Hz * (50 nV/sqrt(Hz)/1.5 nV/sqrt(Hz))^2 ≈ 300 Hz. This is way more modest...

[1] J. A. Barnes et al., "Characterization of Frequency Stability," in IEEE Transactions on Instrumentation and Measurement, vol. IM-20, no. 2, pp. 105-120, May 1971, doi: 10.1109/TIM.1971.5570702.

[dietert1]

1/f noise is just a model like a gaussian error distribution. Technical noise doesn't have those infinite tails.
I suspect that quite often people use the term "1/f noise" as an excuse for deficiencies of their setups and they don't really know what is noise and what are processes caused by ambient conditions, like EMI. The noise level may be much smaller.
When i look at the daily average difference of two voltage references and i see seven consecutive days with 0.01 ppm agreement, i suspect the setup suffers from ambient processes, not from noise. I think it's worth looking at measurements.

Regards, Dieter

[Kleinstein]

A major part of the low frequency noise is popcorn type noise with more or less discrete jumps. This is especially true for the LM399 reference.  This noise part should not go up idefinitely to lower frequencies, but more like have a maximum at the typical rate of the jumps and than less noies at even lower frequencies. This would still be rather low and had to get with an analog filter. Ideally one would also do more than simple linear filtering, but more like identify the upper and lower levels and use a fixed mixture as the reference value, not the ratio as is comes in in that patricular timeframe.

On the issue of settling time it is possible to use smaller resitors to charge the caps and let the DA settle (for still a long time though). These smaller resistors could be switched on and off with analog switches or relays.

For the settling of the DA it does not really matter if the filter is switched to a different mode. The problem are the long internal time constants that than produce an appearend leakage current. The main idea for the LNAs is to keep the input capacitor charged to about the right voltage also when the actual amplifier is not used, maybe a week before use.

One way to reduce the effect of DA is to use low loss capacitors, so ideally PP dielectric, though they are relatively large in size and small in capacity, so that leakage/bias from the amplifiers can become an issue. Motor capacitors are available with some 50 µF for a still moderate price - just a bit bulky. Polyster capacitors may also be OK, though they need about 10 times longer for the DA to settle. The settling time should be about the filter time constant times the Capacitor DA values (with it uncertain way of measurement) divided be the requited accuracy. So for an accuracy target in the ppm range and DA of some 300 ppm for PP capacitors with would be a few 100 time constants waiting time.

The circuits that keep the main filter capacitor at  near zero voltage can reduce the effect of DA due to the smaller and relatively constant voltage. However there is still the settling of the capacitors used for the voltage at the foot point. DA of this 2nd capacitor gives drift in the voltage there and this couples through the main filter cap to the signal. The drif part is still attenuated by the comparable range cross over frequency. So the drift compensation circuit also helps with the DA settling.

[KT88]

On the issue of settling time it is possible to use smaller resitors to charge the caps and let the DA settle (for still a long time though). These smaller resistors could be switched on and off with analog switches or relays.

For the settling of the DA it does not really matter if the filter is switched to a different mode. The problem are the long internal time constants that than produce an appearend leakage current. The main idea for the LNAs is to keep the input capacitor charged to about the right voltage also when the actual amplifier is not used, maybe a week before use.

You missed one important point of the circuit I provided: leakage current caused by DA is provided by the servo after the current doesn't saturate the servo output anymore.
C1 and C2 won't be charged at all and thus won't cause any leakage even by DA...
The faster C3, C5, C6 and C9 are charged the faster the circuit will be settled.

[macaba]

KT88 - nice circuit idea. Perhaps you could add DA-emulation network, in the simulation, on each capacitor?

ADA4523 has a 1/f corner of around 20mHz (actual measurement, not hypothesis) so not much point in having a filter cutoff below that. ADA4523 is still a good choice though, the two other devices that are better have their own issues (ADA4528 max supply too small, OPA189 has anecdotal reports of poor behavior on inputs).

[DavidKo]

Similar approach is used in AN177 from Renesas (page 30). They subtract voltage of low noise reference from DUT. Maybe such an approach can be combined with battery source instead of reference and use pot to tune the voltage to get the necessary voltage to subtract (I have thought about PWM tuning of battery voltage, but I do not know how much noise will be there from PWM modulation - frequency should be >>10Hz, easy filterable, probably will avery in <10Hz).

To get subtraction working man do need low noise reference which can be subtracted from the source. Long term stability will not be so much important. To use the same reference, but inverted you need to shift the signal between each another for a several "periods" before summing to remove the DC signal.

I have thought about using chopper (switch between positive and negative value), make the sum of positive and negative value, you will get the average signal between two following parts, but the signal will be different to noise itself, but for the characterizing it can make sense. Such an approach can induce the chopper noise, period should be short enough compared to the signal change, but not too short to get reasonable signal.

[maat]

A major part of the low frequency noise is popcorn type noise with more or less discrete jumps. This is especially true for the LM399 reference.  This noise part should not go up idefinitely to lower frequencies, but more like have a maximum at the typical rate of the jumps and than less noies at even lower frequencies. This would still be rather low and had to get with an analog filter. Ideally one would also do more than simple linear filtering, but more like identify the upper and lower levels and use a fixed mixture as the reference value, not the ratio as is comes in in that patricular timeframe.

Almost  The PSD is constant at low frequencies and rolls off with 1/f² at high frequencies. The Allan deviation shows a characteristic bump. I have attached a simulation result using a continuous-time Markov chain. I also added the Python code to play with. Put both files in same folder and install the dependencies. Have fun.

Edit: I forgot to mention the following. The plots show different burst noise spectra for varying parameters. τ_1 and τ_0 are the lifetime of the upper (1) and lower (0) state. In those plots τ_0 = 1 s and τ_1 is changed. The maximum noise is found for τ_0 = τ_1, obviously, because this will yield the maximum transitions per unit of time.

[KT88]

@macaba: A crude DA simulation could be quite simple...just add a current source and apply a linear slope. Yes that's not the real deal but it would demonstrate the function of the servo...
One thing I forgot is to bypass C1 and C2 at start-up...
Btw. this servo approach should work for high pass filter as well.

[miro123]

So is there a way to cheat fate? Yes and no. 1/f noise has an Achilles' heel. It is correlated. So in other words, it does give you some information about its future. This can be used by a predictive filter like a Kalman filter

Last month I'm busy with implementation of Kalman Filter on multi-reference board. Unfortunately no time to document or share it. i have learned something from HW V.1 =- I need voltage sources with different properties. (long,shor stability noise TC etc) The Kalman filter work much better - get best of all sources. Some old notes I put them  here https://www.eevblog.com/forum/metrology/lm399adr1399/msg3759236/#msg3759236
I've been playing with supercaps too. Aside the leak and DA they TC is huge. I suspect their capacitance is strongly temperature dependent since C*U=const. As alredy Math siad you best voltage source. I V2 I want to combine the same LM399s and ADR1399 - and maybe  supercaps as popcorn less source to Kalman filter. Hopefully those supercaps does not exhibit any non-linear behavior /e.g. hystreresys etc/ and they fit in simple mathematical model.

[iMo]

A major part of the low frequency noise is popcorn type noise with more or less discrete jumps. This is especially true for the LM399 reference..

What about a "popcorn detector" - something hw-wise simple, counting the "edges" of the jumps. Thus no need to use a DMM. Let us assume the edges of the "jumps" are "fast", something with a

high_pass_filter ->amplifier->comparator.. 


PS: something like this (4uV jumps @7V, buried in some 20+uVpp signal, 1-100Hz):

[Kleinstein]

Chances are it would need 2 comparators to detect the jumps up and down sparate and than some flip-flop to remember the current state.  If this effort is wirth it depends on how good the detection works / how domninant the popcorn noise is.  For references with rather prominent popcorn noise like he LM399 it may be possible to get low enough an error rate to get an improvement and not extra noise from wrongly detected jumps. It also gets more complicated if there are multiple levels, or multiple jumps in one direction possible, so more long lived states involved.

From what I have seen so far the LTZ1000 usually does not show that much popcorn noise and in this case trying to identify jumps may do more harm than good. With my  JFET based reference It looks more like multiple levels (e.g. multiple long lived states) and thus no easy solution.  So it really depends on the reference to start with.

[Gerhard_dk4xp]

In a previous live, 1978 or so, I stumbled across a RCA application note
that did about that.

Cheers, Gerhard

[iMo]

Chances are it would need 2 comparators to detect the jumps up and down sparate and than some flip-flop to remember the current state..

I've added a schematics above - the edges (100ns fast let say) are easy to get off the noise mess, in the first iteration we do not need to detect the current state, but rather to look at a 399 sample - how it "pops". You may divide the number of edges by two and you get the number of pops/time. Thus a simple popcorn indicator..

[Gerhard_dk4xp]

I still have the data book. 1974. I still was in school then.

ed: someone was so nice to scan it:
<       https://www.mikrocontroller.net/attachment/179195/RCA_ICAN6732.pdf         >

[David Hess]

Attached are some photographs of flicker noise that I tracked down inside the analog channel switch of a Tektronix 2230 analog and digital storage oscilloscope.

The best physical model I found for flicker noise is trapped charge with multiple time constants, so identical to popcorn noise, but at a lower level and with multiple time constants all added together.

[mawyatt]

When RCA first introduced the CMOS based Op-Amps way back, many had serious Popcorn noise. We utilized RCA CMOS Logic and Analog right from it's introduction and "experienced" this effect first hand 

This was eventually traced to surface contamination and significantly improved with a "cleaner" process & packaging, and even chip redesigns, and likely initiated the Burst Noice App note mentioned, since RCA CMOS was well known to have this problem!!

Best,

[dietert1]

It's hard to believe the problem hasn't been solved up to now. These stories are ancient and one would assume that semiconductors are near perfect nowadays. Maybe high precision analog is similar to a moon rocket in that one needs to start from the basics over and over.

Regards, Dieter

[iMo]

The 1/f and popcorn problem was solved already..
..Before the Big Bang..

[David Hess]

It's hard to believe the problem hasn't been solved up to now. These stories are ancient and one would assume that semiconductors are near perfect nowadays. Maybe high precision analog is similar to a moon rocket in that one needs to start from the basics over and over.

Popcorn and flicker noise *are* solved problems.  We know what causes them and we know how to minimize them.  Some devices, like MOSFETs, have higher flicker noise because of how they are constructed.  We also know how to remove flicker noise with chopper stabilization.

[iMo]

..Popcorn and flicker noise *are* solved problems.  We know what causes them and we know how to minimize them..

We perhaps know how to "minimize them" but the stuff has not been understood yet. Both exist in any systems you may imagine and most probably are related to fundamental properties of our current universe..

[David Hess]

..Popcorn and flicker noise *are* solved problems.  We know what causes them and we know how to minimize them..

We perhaps know how to "minimize them" but the stuff has not been understood yet. Both exist in any systems you may imagine and most probably are related to fundamental properties of our current universe..

Here we are talking about transistors and PN junctions where charge trapped in defects causes both.

[dietert1]

A nice introduction to 1/f noise is http://www.scholarpedia.org/article/1/f_noise. Must be 2007 or so.
Elsewhere i found that electromigration at bond connections (metal to semiconductor) also generates flicker noise. But all this does not predict the noise level. It isn't enough to observe noise above white noise below a certain corner frequency. That may be any low frequency process and whether it produces 1/f noise over several decades needs to be checked.
One example i have is the observation of an ADR1399 evaluation kit i got recently and prepared well enough for noise levels near/below datasheet specs. But only during the night. During daytime, when the photovoltaic generators 3 m above our lab are working, noise level increases by a factor two - depending on how bright the sun shines. There are no two levels or so. Maybe i need to log the generators, there used to be ethernet connectivity.

Regards, Dieter

[iMo]

As a student at uni I worked in a dept. where they made 1/f noise research (80ties). I built a chamber (a bigger shoe box dimensions), made of a 30mm thick soft iron sheets (welded together), inside walls covered by a copper foil. There was a single hole in it, around 5mm dia, just for the coax (and a low noise preamp inside I built for them as well). Samples and amplifier powered by large stack of batteries placed inside. I can remember the top cover (a top hatch with a handle) was so heavy, that some colleagues had hard time to operate the chamber. I can still remember the sound of the cover when the cover hit the box (the touching surfaces had nice fine mirror finish, thus the box was almost airtight). Luckily the measurements were pretty long so no serious muscular diseases were acquired..   They made subsurface zeners measurements with the chamber as well.