Now as AG7CK writes about it, the schematic indeed looks a lot of the "precision reference" aplication circuit found on ie. old OP05/7 datasheet.. Or from Conrads "oldschool reference" or similar thread.
Did some first measurements and found, that the reference as is is pretty sensitive to humans hand. If I touch the grounded case and thus the connection "A" the output voltage shifts. This sensitivity also represents as what looks like big noise. If the case is untouched the noise is quiet high, but touched and holding hand to it the noise decreases besides a bigger voltage shift. Need to further dig in to it.
-branadic-
After some minor modification (connect "A" as short as possible to GND), the hand sensitivity vanished. I started measurements on that reference and here are the first results.
-branadic-
With the current data given, some dependancy can be plotted:
1. voltage vs. temperature
2. voltage vs. humidity
3. voltage vs. dew point temperature (includes temperature and humidity)
4. voltage vs. ambient pressure
-branadic-
Updated the diagrams above. I wonder why noise is still in the same order of a single LM399 instead of being halved as this is a 4x LM399 reference.
-branadic-
Noise adds as square-root of number of sources only for uncorrelated noise.
So, if noise is not halved that does suggest that there is a correlated noise source dominating the performance.
Temperature?
Another possibility is that one of the sources is much worse than the others. You would need to measure them individually to quantify this.
I had to interrupt the measurements today. The graphs above are updated. Seems like it took quite a while for the reference to acclimate. However, the noise is still bigger than expected.
Measuring each single reference is somewhat complicate and time consuming, so this is a task for later. Next step is to turn back to my other LM399 and to measure the difference between reference and gained output voltage.
Even with first 100.000 measurement values hidden I can't find a correlation to temperature, humidity, dew point temperature or pressure. So it seems that this is just the raw behavior of the reference. Attached Allan diagram without the first 100.000 values and all data.
Any further comment on interpreting the data above is welcome.
-branadic-
The Allan variance plots look like there is quite some 1/f noise, which is no a surprise for the LM399. The 1/f noise can scatter quite a lot - at least it does it with amplifiers. So it could be just one bad LM399 in the set, or maybe a lot of not so good ones. If the unit was not used for a long time, there can be some drift due to humidity effects (board stress) - this can be delayed quite a bit and thus a direct correlation with measured humidity might not be visible.
For measuring single references it should still be OK to use the circuit with 4 refs, just take the signal directly from one reference. There will be a little noise from the other refs. Mixed in but due to the very low impedance of the LM399 this should be only a minute part (e.g. less than 1%).
I don't think that this reference has increased 1/f noise (see FFT plot of all data points). It took the frist 100.000 samples @ 0,25Hz thus around 5 days to stabalize the conditions in the aluminium case. Since that noise is superimposing stability I think. I will investigate each reference output noise in near future.
I'm not sure about the history of the reference, maybe Philipp can leave a comment on this?
-branadic-
The spectrum looks quite a lot like 1/f noise, though the exponent seems to be a little less than 1. So more like E_N = const * f^-0.9.
Having higher 1/f noise would be having the constant factor high - so hard to tell from the spectrum. If at all it would show in the higher frequency part (e.g. 1 Hz - 1 kHz).
I'm not sure about the history of the reference, maybe Philipp can leave a comment on this?
Sorry, I don't have any further informations. The reference was sitting inside a broken Data Precision calibrator instead of a LM399.
The spectrum looks quite a lot like 1/f noise, though the exponent seems to be a little less than 1. So more like E_N = const * f^-0.9.
Having higher 1/f noise would be having the constant factor high - so hard to tell from the spectrum. If at all it would show in the higher frequency part (e.g. 1 Hz - 1 kHz).
Well, comparing the spectrum of this 4xLM399 with the one of my single LM399 at its reference voltage output:
I can't see any difference. And as they are both fully different voltage references with zeners from fully different batches but both based on LM399 I assume that this is a typical spectrum for LMx99 zener references.
-branadic-
Hello,
do I understand it right?
The FFT is calculated from the 100NPLC measurement values?
In this case most of the relevant noise should be filtered out.
I would rather compare 1/f (0.1 - 10 Hz) and eventually wideband (10 Hz - 100kHz) noise.
On the 1/f noise I usually find a large stray between different LM399.
See difference between my LM399#1 and LM399#2 with 1uV/Div and 1sec/div
with best regards
Andreas
Hello,
do I understand it right?
The FFT is calculated from the 100NPLC measurement values?
In this case most of the relevant noise should be filtered out.
That's right, but as far as I understood you can't filter 1/f noise? How does that fit togehter?
-branadic-
For comparing noise data, the 0.1 Hz to 10 Hz band is a standard frequency range. Using 100 PLC mode of the DMM will filter out noise above about 0.25 Hz as it should only give one value every 4 seconds. So there is not much left from the more standard 0.1-10 Hz band. It is still OK to use such slow sampling, but one gets different data and thus difficult to compare too others who are less patient.
To get closer to the normal 0.1 -10 Hz Band it would need the 1 PLC speed setting.
If a DMM is used for the measurement, I would suggest using 1 PLC speed and the Allan variance curve to judge noise. It has kind of similar information as the FFT. The very long time scale might add some extra noise from the DMM due to effects like thermal variations - thus not much sense in using a 100 PLC mode, unless the DMM is old and does not perform well faster.
Andreas,
have you tried using your Picoscope plus 0,1 - 10Hz noise amplifier to plot spectrum?
Kleinstein,
wouldn't we expect noise below 0,1Hz to still have a negative slope like in the 0,1 - 10Hz band? Do we know how this slope changes for very low frequencies? Are there any references or publications available?
-branadic-
That's right, but as far as I understood you can't filter 1/f noise? How does that fit togehter?
Hello,
at least it is very difficult to set up a analog filter below 10 Hz which does not introduce additional noise in this frequency range.
The integrator in the DMM in a multislope ADC works different: part of the integrator is in the digital domain which does not add additional noise.
have you tried using your Picoscope plus 0,1 - 10Hz noise amplifier to plot spectrum?
not up to now and I have not much measurements from LM399 with PicoScope.
When I take a old measurement from LM399#CH6 with my usual setpup I get this:
But this is mainly the filter characteristic with 10Hz upper bandwidth.
I guess for a FFT I would need a different setup to see the lower frequency parts.
with best regards
Andreas
The AC coupling could fool you when doing the measurement with the scop. So it could be compensator for the 1/f slope of the noise.
I would rather compare 1/f (0.1 - 10 Hz) and eventually wideband (10 Hz - 100kHz) noise.
On the 1/f noise I usually find a large stray between different LM399.
Did that today and measured low frequency noise with Pipelines amplifier. Here is the result. The gain should be A=10.000 so one vertical division corresponds to 500nV.
-branadic-
I noticed in the scope photo that it was AC coupled during the measurement. The Tek2465 has a 10Hz (-3db) lower frequency limit when AC coupled.
With the 0.1-10Hz noise filters/preamp the scope needs to be DC coupled. The noise filters/preamp has an internal AC coupling that works down to 0.1Hz.
I expect this combination of equipment and settings provided you with a tuned filter at 10Hz that was -6dB down from the actual 10Hz noise.
Thanks for the hint chuckb, I corrected that.
-branadic-