DMM Noise comparison testing project

[Kleinstein]

The Alam variance curve looks a bit odd, especially for the short times, below some 2 seconds. The values there are surprisingly low. This makes me think the first values could be "wrong" and not the maximum at some 2 seconds. With "wrong" I mean an issue in the meter that causes correlation between the readings (e.g. a hidden rolling average, possibly just for the zero readings).

[dietert1]

I thought about it, too. Of course there could be some cheating inside the K2700, who knows.
When debugging the fiber ethernet interface i saw a K2700 timing pattern that shows an extra cycle every eight readings, so about once per second. Probably the autozero. Yet another additional cycle happens every 10 seconds. Probably an autocal type adjustment using the LM399 reference. A 5 PLC reading with line sync becomes 120 msec. The observed reading delays are 120 msec, 240 msec and 360 msec.
In my understanding the "flat" part of the Allen diagram above one second is due to autozero. Maybe i can repeat the test without autozero.

Regards, Dieter

[zrq]

This is the noise Allan variance of L4411A (UI less version of Agilent 34410A/11A) with input shorted, AZ ON, and set to 1PLC. Plotted either in volts or ppm of range.
It have a very low noise in the 10 V range that easily reaches 0.1 ppm of full scale at 0.1s averaging time and <0.02 ppm in 2s (close to 100 PLC), but the 1V and 100V ranges have slightly lower noise compared to the older generation 34401A.

[Kleinstein]

I thought about it, too. Of course there could be some cheating inside the K2700, who knows.
When debugging the fiber ethernet interface i saw a K2700 timing pattern that shows an extra cycle every eight readings, so about once per second. Probably the autozero. Yet another additional cycle happens every 10 seconds. Probably an autocal type adjustment using the LM399 reference. A 5 PLC reading with line sync becomes 120 msec. The observed reading delays are 120 msec, 240 msec and 360 msec.
In my understanding the "flat" part of the Allen diagram above one second is due to autozero. Maybe i can repeat the test without autozero.

Regards, Dieter

The classic AZ mode should suppress 1/f noise but add white noise: 2 readings and thus 1.4 times the noise and half the reading rate.
The resulting white noise should give an slope of -0.5 in the log-log Allan variance curve. A flat part and than going up is from additional low frequency noise, e.g. thermal fluctuations / drift in the front end and protection.

Getting the different reading times can be confusing and points to a not so simple AZ mode, like doing a zero less often and maybe additional averaging for the zero. For the Allan variance this is a bit "cheating" for the shorter time scale and thus a better result there. It however comes with more noise on the slightly longer times. The more complicated AZ mode can make some sense in some applications, but it would be nice to also have the simple, classic version as an option as a constant reading rate can also be a big plus.
A non AZ test could also make sense. One could kind of simulate the classic AZ case with these data, though withough the switching transients.

The L4411A noise looks very white with very little extra low frequency noise and drift.

[dietert1]

I thought about it, too. Of course there could be some cheating inside the K2700, who knows.
..

The user manual describes data processing and its configuration in great detail. My Allen diagram happens to document the K2700 behavior in its default configuration after power up (filter on, window = 0.1 %, 10 readings, moving average). One can turn off the filter with the expected effect on noise at averaging times less than 3 seconds.
I also verified that the observed reading delays are gone when autozero is disabled. Rate becomes a steady 120 ms per reading. In the manual they mention that their autozero employs a moving average and needs some time to settle.
Remains to check the effect of temperature at high integration times.

Regards, Dieter

[Alex Nikitin]

Here are some results for the HIOKI DM7275 voltmeter, shorted input, NPLC1 with 50ms intervals and NPLC10 with 500ms intervals, 1hour each run, 100mV, 1V, 10V and 100V. The vertical scale in Volts, obviously.

Cheers

Alex

[dietert1]

Here i have another Allen diagram for the K2700 as nullmeter, this time without the K2700 filter and with a temperature linear drift of 21.7 nV/K subtracted. A SHT45 sensor connected to the fiber ethernet device was used to log the internal temperature inside the K2700 module bay together with the meter readings. The drift subtraction improves results beyond 300 seconds integration time. The ultimate resolution of the K2700 as nullmeter is better than 10 nV.
Later there may be more data, yet currently i need the setup for development work.

Regards, Dieter

[ballsystemlord]

So all of the submitted data can be found here: https://xdevs.com/datashort/ but how do you use it? Like, I thought that we'd have a graph, bar chart, or histogram of the info, but I cannot find one.

Thanks

[Kleinstein]

There are a lot of raw data in the folder. Raw data a nice, but really lengthy. One step more processing to RMS noise or specific points of the Alan variance curve would be nice, but I have not found them.
To use the raw data one would do some processing on the data:
1) make shure there is no significan drift, and if needed maybe remove it
2) a) calculate the RMS noise for a suiteable sample length
    b) calculate the Alan deviaton (needs relatively long sets) and look at a few characteristic points
One could than create the sets for different meters, points of the noise data

Ideally (white noise) one would get away with just looking at the RMS noise for something like 1000 readings. However some meters (especially Keithley) have non white noise as seen from an odd looking Alan var curve. Because of this the RMS noise can depend quite a bit on the test set length.
There is also a small difference between individual meters of the same type and 50/60 Hz can make a difference.

Inside this thread there are a few graphs and also some processed data (e.g. RMS noise for a few cases).  One difficulty is a comparison with different lengths for the data sets and also a few points that can be off (errous scaling (ppm/FS), odd settings like digital filter active or with significant background drift).