Low cost LM399 and ADR1399 noise indicator

[iMo]

Below the V3, 78L05 replaced by the TL431 set to 5V.
No instability even with the 1000uF at its output (here one has to be careful with the 431), voltage 5.00V, no change in 3 digits when sinking 20mA.
ADR1399 still shows 2.2-2.6uVpp..

PS: finally replaced the 7815 by 78L15.. Works fine.

[iMo]

Now, I want to get an indication of the noise level and noise type upon a set of my 399/1399s..
What would be the recommended process here ?

For example:
Procedure 1.
1. place a new sample and close the box
2. wait 20 minutes until it settles
3. make a single shot indication (10secs, peak detect mode, etc.) and record the mVpp/10 = sample_uVpp
4. goto 1.

Or
Procedure 2.
1. place a new sample and close the box
2. wait 20 minutes until it settles
3. make, say, 10 single shot indications at random times  (10secs each, peak detect mode, etc.) and make the average_mVpp/10 = sample_uVpp
4. goto 1.

Or
Procedure 3.
1. place a new sample and close the box
2. wait 20 minutes until it settles
3. make, say, 10 single shot indications at random times (10secs each, peak detect mode, etc.) and write down the worst (or best, or both) mVpp/10 = sample_ uVpp
4. goto 1.

or..

Any recommendation before I start to mess with my indications ?

[Kleinstein]

The best results are probably from the 2nd procedure, so the average of the peak to peak values. Ideally one would remove something like the best and worst of the set before taking the average to avoid rare excursions, though this is somewhat debatable.
Anyway the noise is usually more complex than just the 0.1 Hz to 10 Hz band. This band is more chosen because it is reasonable easy to measure (e.g. such an amplifier), not so much because it is relevant for real world use. For a DMM or so it is more the noise for lower frequencies that is relevant and audio is obviously higher.

[iMo]

The best results are probably from the 2nd procedure, so the average of the peak to peak values. Ideally one would remove something like the best and worst of the set before taking the average to avoid rare excursions, though this is somewhat debatable.
Anyway the noise is usually more complex than just the 0.1 Hz to 10 Hz band. This band is more chosen because it is reasonable easy to measure (e.g. such an amplifier), not so much because it is relevant for real world use. For a DMM or so it is more the noise for lower frequencies that is relevant and audio is obviously higher.

The 2nd procedure introduces a kind of addtional filtering, imho.
My naive preference would be 3. with recording the best and worst for a given sample.. But not sure, of course.
I wonder how the big boys are doing it in their datasheets..

[Andreas]

Hello iMo,

what I am doing:

1. place a new sample + cover with a micro fibre chloth (avoid local air movements)
2. wait some time until it settles (also the amplifier).
3. make 15-19 successive single shots (minimum 10 secs each, but mostly 100 secs per shot)
    settings: 20 MHz hardware BW limitation,  minimum 50 kSps/sec, 1 kHz low pass software filtering to filter out oscilloscope noise (it is far outside 10 Hz).
4. build average and standard deviation of uVpp and uVrms values
5. also check for "popcorn" events in the records. (here zoomed in record 16 of 19)

Of course the 100 sec measurements deliver somewhat "higher" averages than 10 sec measurements.

with best regards

Andreas

[Andreas]

I wonder how the big boys are doing it in their datasheets..

There are application notes:

(I think it was LT) with following procedure:
make several (hundreds) single measurements of 10 seconds.
sort the results by peak values.
take the value where half of the results are above and below the final result (so effective the "median" value).

with best regards

Andreas

[iMo]

My Rigol's DS-1062CA (I would call it an "electrician scope" ) setting is currently:
DC, Acquisition: peak-detect mode (my understanding is it emphasis showing peak values), sinx/x OFF, Sampling rate - no idea, BW limit ON, time base 1sec/div (it got 12secs over the display), printing Vpp on the display in addition.

The median over hundreds of measurements is a pretty heavy low-pass filtering.. Almost any sample after such filtering will be better than the competitor's best

[Andreas]

But it is only a "typical" value (see e.g. Data sheet of LTC6655 for the procedure)

with best regards

Andreas

[iMo]

Ok, interesting method with the 6655..

[iMo]

After cleaning up the boards with IPA, final assembly and installing foam insulation (from the bottom of the pcb and a cap for the samples) here are the first data with the brief notes..

PS: added a shot of the foam insulation used - there is a small block of foam beneath the pcb as well.

[iMo]

And the results after a weekend with the noisy refs..
For example:
..the LM399H will have equal or less than 4 to 5.5uVpp noise in 80% of 10secs intervals..
..the ADR1399 will have equal or less than 2 to 3uVpp noise in 80% of 10secs intervals..
See may above table - the top green one is the number..
 

Some notes:
- almost all samples do "pops" from time to time (one 399AH does nice ones)
- no Morse telegraphing this time - the RC snubber did help obviously with the "Morse MAC199" and also with ADR1399 - thus my previous observation was an instability perhaps, imho
- MAC199 - all were shortly produced around 1990 by TESLA afaik (TESLA corp. died after 1990), almost unseen (the MAB399 was/is more less known), my measurements from past showed large drift (none snubber at that time, btw), and one of the samples above has broken heater. Interestingly all show 0.1-10Hz noise comparable with the LM399.. 

For a calibration the 1N4148 is suitable (~0.7uVpp here), with an empty socket you will see the noise of the voltage regulator(s), here I see around 350-400mVpp that is 35-40uVpp (and multiply it by 2k7/1k5 divider, say 3x).

[iMo]

FYI - a DIY "ref diode" made of a 6.2V BSX zener and an 1N4148 diode in series.
The noise seen is typically between 1.8 and 2.5uVpp at 3.6mA.

PS: added a 10 minutes long run - 3uVpp