USA Cal Club: Round 2

[rhb]

Yeah, I found that after I posted.  I modified my code, but didn't modify the post.

It's happily logging the FX reference that TiN built.  I won't have 24 hours of data until bedtime, so I'll log the 34401As until tomorrow morning and then switch to the 3478As.

An interesting thought occurred to me as I crawled into bed.  I'm collecting measurements from a single device with two meters over the same time frame.

As some will recall from when I first got active in the forum, I spent my career in  DSP.  I can separate the noise characteristics of the FX and the LM399s in the 2 meters  It's not particularly exotic math, so it's likely that TiN and Andreas are already doing this, though I've not seen a discussion of it.

Briefly, because the 1/f noise is uncorrelated, if one computes the cross correlation of the two meters one gets the autocorrelation (time domain amplitude spectrum) of the FX.  One can then estimate  the autocorrelation of the  1/f noise of each meter by dividing the autocorrelaton of each of the  meter readings by the cross correlation of the meter readings.

If this has been discussed, please post a link so I can post my data there.  If not, then I'll start a thread once i have collected the data.

Edit:  I realized I don't need the cal kit to do the 3478As as I've got 3x LM399s I can use as transfer standards from the 34401As.  So this evening I'll switch to logging the PX reference with the 34401As and collect 24 hours of data on that.  Then I'll do short runs on the resistors and have the kit in the post Thursday which will have it back to @vindoline Saturday with luck and Monday at worst.

[RandallMcRee]

. . .
Briefly, because the 1/f noise is uncorrelated, if one computes the cross correlation of the two meters one gets the autocorrelation (time domain amplitude spectrum) of the FX.  One can then estimate  the autocorrelation of the  1/f noise of each meter by dividing the autocorrelaton of each of the  meter readings by the cross correlation of the meter readings.

If this has been discussed, please post a link so I can post my data there.  If not, then I'll start a thread once i have collected the data.

Edit:  I realized I don't need the cal kit to do the 3478As as I've got 3x LM399s I can use as transfer standards from the 34401As.  So this evening I'll switch to logging the PX reference with the 34401As and collect 24 hours of data on that.  Then I'll do short runs on the resistors and have the kit in the post Thursday which will have it back to @vindoline Saturday with luck and Monday at worst.

I'm interested also in the autocorrelation aspects of this. I don't know of any discussions on this, personally. I have been studying this thesis (hopefully I can attach to this post). It's called "High-Sensitivity Instrumentation for Low Frequency Noise Measurements” by:Gino Giusi https://pdfs.semanticscholar.org/ed82/b314dec4644069fb6a5f7c0bba84465405e9.pdf. Is this the kind of processing you are talking about? I'm not a DSP guy, just fyi, but I can test and am trying to come up with a DIY meter for noise measurements.

I'm next in line--take your time (as vindoline mentioned).

Randall

[rhb]

Cool!!!  Thanks for the thesis!  I need to pull my duplexing printer off theshelf and get to where I can work on it and see if I can fix it.  I much prefer sitting in a recliner with printed paper for doing serious reading.

FWIW the canonical reference on such things is "Random Data" by Bendat & Piersol.  I bought the 2nd ed in grad school 35 years ago which I almost wore out.  I also have the 3rd and 4th which is the last as Alan Piersol passed away.  While the latter is preferable, books have become depressingly expensive and the 2nd or 3rd ed is more than good enough.

It is wall to wall calculus all of which originated with "The Extrapolation, Interpolation and Smoothing of Stationary Time Series" by Norbert Wiener.  That originally appeared as a classified document during WW II and was referred to as "the yellow peril"  by the EEs who had to read it.  The name came from the color of the covers denoting it's classification level.  Wiener wrote well.  I was recently sent a copy of his autobiography which should be great fun.  So far I've  just read the part about the Geophysical Analysis Group which Wiener set up at MIT after the war which pioneered DSP.

Fortunately, like much of mathematics, you don't have to dig very deeply to make use of it.  Most of the agony is in the proof of why it works.  Though there are always rare instances when the fine print is *very* important.

I powered up the PX reference to let it warm up and one of my 3478As reads *exactly* 7.0456 V.  The other reads 7.0458 V.  I was reading the specs on the 34401A last night and both of mine are well within the 1 year accuracy interval.  The nicer looking one was a closet queen and has never been calibrated since it left the factory.  I got the 3478As for ~$110 each and the 34401As for ~$250 each.

Looking at the really nice graph TiN put on the FX reference makes me want to investigate the thermoelastoplastic deformation of silicon and the construction of the LTZ1000.  I think a fairly simple model would account for the hysteresis seen when it is cycled over the 20-50 C temperature range.  Over the 33-50 C range it appears to be purely elastic without hysteresis.  Then the heating and cooling curves diverge until crossing again at 22-23 C.  Differential expansion between silicon and gold bond wires might well account for the differences and it should not be particularly difficult to calculate the stresses and elastic-plastic threshold.

Have Fun!
Reg

[bitseeker]

Oddly enough I was just reading about this in the manual.  I think resolution is a measurement of the smallest size.  Here is what it says:

"Specify the resolution in the same units as the measurement function, not in number of digits. For example, for dc volts, specify the resolution in volts. "

and provides the example:

MEAS:CURR:AC? 1,1E-6                      6-1⁄2 digits on the 1 A range

I think you also need to set the range explicitly.  This sets both the range and resolution:

Sense:Voltage:DC:Range 10, MIN

In my logger, I first reset the 34401A (and 34410A) in order to have a known starting configuration. Then, set up the range and other parameters that differ from the starting config.

i.e.,

  *RST; *CLS

followed by

  CONF:VOLT:DC 10,MIN; :INP:IMP:AUTO ON

[rhb]

I just plotted the logs for 22.5 hrs with my 34401As reading the FX reference.  The old, very heavily used meter showed about 1 ppm total variation over that period with no visual correlation with temperature.

The nice shiny closet queen showed 2 ppm variation with strong correlation with temperature.  I've decided to continue logging the FX until morning.  That will pick up the 1-2 degree drop overnight.

I had been planning on sending the kit back out on Thursday, but this result is very interesting.  How do other people on the waiting list feel about my keeping it a little  longer and shipping it out Saturday instead?

I'll make plots and run some statistical analysis in the morning and post the results.

Reg

[TiN]

Upload some CSVs to FTP, so others can have fun too. 
Instructions.

[rhb]

Columns are:

seconds since start of run
temperature C
relative humidity
heavily used 34401A
closet queen 34401A

All data are from the FX reference.  The spike in the humidity was my doing a load of laundry in the washer and dryer 10 ft away.  There is a 6000 BTU window A/C cooling the 7' x 10 ' closet which is my current lab area.  Air is deflected towards the ceiling as much as possible.  There are 2x 34401A, 2x 3478A , 3x Z400 and a linear PSU running in the room which has the door open to the rest of the house.  My presence in the room is very noticeable.

I'll post a longer copy tomorrow.  I'm only doing it now because I woke up at 2 am and an hour later am still awake :-(

I'll upload to xdevs when I've finished the run.

FWIW the heavily used 34401A was wonky when I got it from ebay and did not produce any usable data.  I "fixed" it by cleaning a 1 cm splotch of flux residue from where an LF357 had been replaced.

Edit: Data have been removed as the full run set is posted later with FX, PX and ratio in a single file

[Grandchuck]

  How do other people on the waiting list feel about my keeping it a little  longer and shipping it out Saturday instead?



Reg

What you learn could help others on the list.  Take your time.

[rhb]

I have turned off the A/C, so the room will warm up to 85 - 90 F during the day as there are 3x HP Z400s running to provide heat.  I'll then turn it back on around 4:30 to finish off the FX run.

What I'd like to do is repeat the same run with the PX followed by doing a similar PX/FX ratio run Saturday and Sunday and ship the kit back on Monday.  That will give me 3 series of ~100,000 samples spanning 48 hours and including a ~20 F temperature excursion.  Not as good as an environmental chamber, but the best I can do as that build is only partially completed and is not large enough to hold the meters anyway.

I've been making a list of data plots to generate and those 3 data series will keep me busy for quite a while.  Each of the data series will involve making about 20 plots and possibly double that.  I have a suite of statistical tools I wrote when I was analyzing rock properties from several hundred GB of data from oil and gas wells, NOAA and other sources.

Edit:  I've attached a plot showing the distribution of errors based on the nominal 10 V output from the FX.  It's just part of the data, but I wanted to see what the distribution looked like.  This is strictly just an initial peek at the error distribution.  I'll repeat later for various temperature ranges.  I had to modify my program for computing the histogram to handle doubles.  Rock measurement errors are 1-2% *if* you're lucky.

[rhb]

I made new plots of the data and have attached the input file to this post.  For some reason I got an argument from the server about saving it.

This is all just exploratory data analysis.

[TiN]

Some plots from rhb data.

[rhb]

I can't read the text on the plots.  Would you please use a bigger font?  And also explain what you're trying to show.  The "closet queen" (col 5)  has much more temperature dependence than the "work horse" (col 4).  Your plots appear to show the opposite, but even using a magnifying glass I can't be sure about the curve keys.

I posted what I did just to give a general idea of what I was doing.  in about 20 minutes I'll switch to the PX reference, at which time I'll post the full data series file for the 48 hour FX run.    I'm not sure when I'll have time to do plots and other analysis as I've had personal matters intrude upon my having fun. :-(  But taking care of family takes precedence over playing.

[TiN]

I don't see any issue with size, did you download large picture?
My numbering is from zero, so col3 (green) is more stable meter, col4 (blue) is less stable one.
Otherwise everything should be straight-forward. One plot in timescale, second is temperature (ambient) dependency.

[rhb]

col 1: time in seconds since start of run
col 2: temperature in Celcius
col 3: relative humidity
col 4 work horse voltage
col 5 closet queen voltage

My 1600x1200  screen is in portrait mode which makes the text slightly smaller, but it's still too small even if you use landscape mode.

I made my figures to professional society meeting presentation standards.  Until I retired shortly before employment for very expensive old guys went away in 2008, I had missed 3 annual Society of Exploration Geophysicists meetings in 28 years.

I used gnuplot with "set term post color 'helvetica-bold' 14 lw 3".  I *think* that's readable in a hall with 500 people in a normal projector setup.  I've never had a chance to put one up and walk to the back of the room to check though.  A size 16 font might be needed in a room that large.   My sub-specialty had smaller rooms assigned that only seated 200.

That may seem like a minor detail, but if you fail to communicate with the audience it's a waste of everyone's time.  I've seen far too many grpahics which were impossible to read even from the first 10 rows.  The only justification for enduring it was you wanted to see the next talk.

Real life has intruded into my play time, so there may be some delay before I can post more analysis.  I'll post data as it becomes available and ship the kit to @vindoline on Monday,

I'll probably spend 100+ hours on analyzing the FX, PX & PX/FX data.  There are a lot of things to look at in that data set.  So results will trickle out as time permits.

[rhb]

OK Here's the full 48 hour FX run.  The PX is logging and I'm off to bed.

Many thanks to to all who made this possible, but especially  @cellularmitosis, @WaveyDipole, @ vindoline and @TiN.

Have Fun!
Reg

Edit: Data have been removed as the full run set is posted later with FX, PX and ratio in a single file

[TiN]

That may seem like a minor detail, but if you fail to communicate with the audience it's a waste of everyone's time.

Image has decent enough resolution and size, so that is not an issue of the generated data.
If it is open and resized afterwards to tiny phone screen or obscure 1200x1600 resolution - how is that my problem? After all you can kill some tree and print on A4, where it will be all readable. 
Since this is goodwill I'm not investing lot of time for best possible formatting, but rather reuse my existing plotter that was made for much bigger data set (18 DMMs) with your data. Hence the smaller font-size.

[bitseeker]

Auto-scaling will make it small. View the plots at 100%. They're nice and big.

[MaxFrister]

Tin: Thank you for the very cool graphs.  Can you say a bit about what they represent.  I assume the shadow colors are the samples and the dark lines are some form of trend or average.  How were they made?

Rhb: Here is what the legends say (you will need to click on the picture and then save it to view it full size). The originals are large enough resolution that you can enlarge them as well.

[MaxFrister]

For the next person setting up logging for a 34401A:

One other important hint that bitseeker gave me is that the 100NPLC read takes approximately 3 or 4 seconds.  I had trouble with pyvisa until I increased the timeout.

After I got it working, I changed from "read?" to "initiate" and "fetch?" so that I could communicate with other instruments while waiting for the 34401A to finish.

The final bit of 34401A SCPI is "display off" to save wear on the VFD display.

BTW, what is a good interval between samples?  As fast as possible?  10 seconds? 1 Minute?  I know some temperature sensors self-heat if sampled too frequently. 

Next up, 34970A...

[rhb]

Over the course of my  career I spent approximately 2000 hours in professional scientific meetings  looking at around a million plots of almost every imaginable type of data.  I offered a suggestion on improving the presentation of  data. 

How you present your data matters a lot.  If your audience can't understand what your figure is showing in a few seconds they will probably not understand your presentation.  Large complex datasets are very difficult to present well.  Doing it well is one of the things that  distinguishes the A list.  If someone has to ask you to explain your figure, it needs work.  And if you want to make the A list, both your writing and your graphics must make the grade.  It is a lot of work.  So your have to care about meeting A list standards to do it.

In the past I have not bothered to comment on poorly done figures posted here, but in this case, as I had trouble recognizing my own data, I thought I should.

I commend to everyone's attention the work of Edward Tufte on data visualization.

https://www.edwardtufte.com/tufte/

Much otherwise excellent work has gone unnoticed because of poor presentation.

Reg

[TiN]

How you present your data matters a lot. 

I have complete agreement on that matter. Axis without labels or mismatched colors haunt me at night, because you forgot what chart shows next week already. 
I only disagree about previous statement that graph is too small, which is not true, I always spend extra effort and storage space to generate higher resolution images.

MaxFrister
Your assumptions are correct. Dark lines are result of gaussian_filter1d from python lib over data, just to see a data trend.
All calculations, like median, min/max, sdev are done on raw unfiltered data, trend used only for graph purposes.
Bottom graph is traditional seconds timescale, imagine just as oscilloscope, but with very slow time/division. I have posted Python plotter generator app sourcecode few posts back, when we went thru bitseeker's dataset.
Top graph is relation between temperature (ambient sensor on X axis) and measurand (Y axis). This graph is not very clear, because temperature change was too little.
Usually better data have temperature swing >5c (e.g. AC on / AC off cycle).

Perhaps rhb could log FX with large temperature swing (e.g. one day AC on, one day AC off) if he is interested in temperature stability of his 34401A.
Because measured FX reference have TC <0.05ppm/K, it's own error usually only a tiny fraction of the 34401A's own TC error.

[rhb]

God says, "No.".  The daily highs were in the mid to upper 90's just before I got logging working. but now it's low 80's and will be until I send it out Monday.  It's nice, so I'm not complaining.

I've got a good cal on the 34401As and should be able to transfer that to my other instruments with careful thought and a few additional gadgets.  The "Work Horse" is more than adequate for testing my voltage references. 

Cellularmitosis offered me the "Work Horse" in an email one morning after he sniped it on ebay on an "if it doesn't work I'll take it if you pay the extra shipping" basis.  Talk about an offer you can't refuse.  I'd seen it, but was unwilling to bid on it because it looked to be wonky, which it was when I got it.  It just produced random digits measuring a DC voltage source.

I opened it up to take a look around and saw a 1 cm splotch of flux around an LF357 which had been replaced.  A series of VCR failures while living in Houston had suggested to me that the hygroscopic nature of rosin was a significant issue in humid environments.  I had already verified that was the case with a half dozen other "repairs" which consisted simply of cleaning residue from consumer gear. So I took a spray bottle of 91% isopropyl and an old toothbrush and cleaned the residue from the board.  It immediately worked like a champ.  But until now I did not know just how good it was.

It has redoubled my interest in investigating the aging and thermal hysteresis of voltage references.  I am certain that very fine work has been done at HPAK, Fluke and Keithley, but it's all trade secret stuff. Unfortunately, I have to have many years of data for such work.  So even if I start collecting data, it may never get finished.  My requests for existing historical data have been rebuffed.  So a few weeks of analysis must wait on several years of data collection and the burden of building and maintaining a data acquisition system for the task. Prior to learning I was next on the list, all my attention was on learning to solve the strains of arbitrary prisms under combined torsion and bending moments.  A subject to which I shall return as soon as the kit is in the post and I have set up logging on some LM399 references.

It took me 3 years of intensive study of 3000 pages to master the  mathematics to solve the problem of forecasting aging.  What are the chances that someone building and logging references as a hobby is going to put that much effort into the mathematics?.

That effort was without question the most challenging intellectual effort I have ever undertaken and I probably have over 30 semester hours of university level mathematics.  But none of it even came close to "A Mathematical Introduction to Compressive Sensing" by Foucart and Rauhut.

And to setup the equations, you have to be able to write out the thermovisocelastoplastic constitutive equations for the wirebond to the reference die.  For starters.  It probably gets worse.  And having dealt professionally in the oil industry with much simpler problems in the mechanics of materials, I don't think there is a snowball's chance in hell that anyone has much chance to do all that on their own.  Humans don't live that long.

So without cooperative effort, it won't happen. At least in my lifetime.

Reg

[tomato]

Over the course of my  career I spent approximately 2000 hours in professional scientific meetings  looking at around a million plots of almost every imaginable type of data.

That's truly impressive.  Assuming each talk at those meetings was 12 minutes (followed by 3 minutes of questions) that works out to a data plot being presented every 5.76 seconds.  Every 12 minute talk would contain (on average) 125 data plots.  I don't know how anyone would be able to absorb any information from that many data plots presented at that rate. 

[bitseeker]

One other important hint that bitseeker gave me is that the 100NPLC read takes approximately 3 or 4 seconds.  I had trouble with pyvisa until I increased the timeout.

Ah, yes. I forgot that I had also increased the PyVISA timeout. I'm glad my mention of the time required was enough to point you in the right direction.

After I got it working, I changed from "read?" to "initiate" and "fetch?" so that I could communicate with other instruments while waiting for the 34401A to finish.

What I did was send "read?" to both meters, but not fetch right away. While the meters were integrating, I then fetched the temp and humidity. Finally, I read back the DMM's values. It seemed to work effectively the same as using initiate and fetch. There seemed to be some overlapping functionality and behavior in the SCPI command set.

The final bit of 34401A SCPI is "display off" to save wear on the VFD display.

Yep. Same here.

BTW, what is a good interval between samples?  As fast as possible?  10 seconds? 1 Minute?  I know some temperature sensors self-heat if sampled too frequently. 

I took samples as fast as the DMMs would let me, fetching temp and humidity at the same rate. So, that'd be around a sample every 4 seconds.

The firmware on the Uno takes 7 samples and averages them every time you request a value, so if there is a self-heating issue, I'm not sure what interval might be problematic.

[rhb]

I've got NPLC 100 set and am getting samples from both meters at ~2 second intervals.  For simplicity I just used the 1 second granularity of the Unix epoch.

I left the displays on so I could easily see what was going on.

Right now I'm as nervous as a cat in a room full of rocking chairs, but once the ratio run is finished and the kit is in the post I'll start looking into the various details.  The Si7021 that Jason used for the Tempduino is very low power, so I don't think self heating is an issue.  I plan to investigate adding an MSP430 based transit logger to the kit as soon as I can get the parts.  As that will be battery powered, sampling at 1-10 minute intervals is probably desirable.

From a statistical perspective, the larger the number of samples the better.  If you plot a histogram of a Gaussian distribution, it doesn't really look Gaussian until you get to around 100,000 samples.

TiN requested photos of the lab setups, so I've attached a couple, one from the PX run and another from the ratio run.

Have Fun!
Reg