USA calibration club

[CatalinaWOW]

My data is now copied into TiNs repository.  I will continue to post graphs of conclusions here, but will leave the raw data elsewhere.

Thanks for the WinSCP suggestion.   Made the operation simple.  Which has a downside.  One more area where the motivation to learn has been eroded.   Years of mandatory exposure to Windows in the corporate world has allowed me to be deficient in Linux and command line operations.  I am slowly switching to Linux for all of the usual reasons, but still find most tasks far easier in the system with decades of training behind me.

On that note I have used Libre Calc and Excel for data reduction.  Neither is particularly good.  Libre Calc is very slow on large files, particularly graphs.  Excel is far faster, but crashes periodically.  Apparently has a memory leak somewhere which catches up on large files.  On to try Gnumeric.  Too bad there is no single tool set that is ideal for everything.

[CatalinaWOW]

Catalina,

Can You please post the calculated temperature coefficients of the three instruments based on the text file.

0.007131491465747643   
9.653462320521867E-05   
0.00020292057647497002

?

Room ambient temperature fluctuations were not sufficient to get a meaningful temperature coefficient for the 3478 and the 3456.  The quantization was too large with respect to the noise and any changes due to temperature.  I am still trying to tease out a temperature coefficient for the 3457, but don't have an answer yet that I consider meaningful. There does appear to be a very slight negative coefficient, but it is a fraction of a part per million per deg at most, and I am not yet ready to say that anything I have seen is statistically significant.

A simple inversion of the test on the SVR-T isn't in the cards because the thermal time constant of the meters cannot be assumed to be short relative to the overnight temperature changes.  I had evidence that the thermal time constant of the SVR-T is in the order of 10 minutes so the overnight temperature changes were tracked reasonably.  From my experience with other somewhat similar pieces of equipment I would expect the thermal time constant of the 3457 to be more on the order of an hour.

Based on the small size of the tempco, I have put any further investigations of this on the back burner while I chase other things of more impact and interest to me.

[bitseeker]

Thanks for the WinSCP suggestion.   Made the operation simple.  Which has a downside.  One more area where the motivation to learn has been eroded.   Years of mandatory exposure to Windows in the corporate world has allowed me to be deficient in Linux and command line operations.  I am slowly switching to Linux for all of the usual reasons, but still find most tasks far easier in the system with decades of training behind me.

Glad WinSCP worked OK for you. Although I have a dedicated Linux box (or two) at home, on my Windows and Mac systems I also have Linux virtual machines. The Mac also has a Windows virtual machine. So, it's easy to switch to whichever environment works best for a particular task or project. Sometimes, though, it can be a bit of a brain juggle getting the fingers to use the correct hotkeys in the correct OS.

On that note I have used Libre Calc and Excel for data reduction.  Neither is particularly good.  Libre Calc is very slow on large files, particularly graphs.  Excel is far faster, but crashes periodically.  Apparently has a memory leak somewhere which catches up on large files.  On to try Gnumeric.

I also tend to use Excel more than Calc since I often work with files on the order of 100K to 1M rows. However, despite the additional speed, I've occasionally had Excel crunching all night (8 threads at 100%) and still not finish. Need mo' pow-ah!

Too bad there is no single tool set that is ideal for everything.

Yep. C'est la vie. I've even used one tool solely as an intermediary to get data into a format that will load into another tool. As the old saying goes, the great thing about standards is that there are so many to choose from.

[alm]

For serious data analysis (1M rows is no longer some casual experimentation), I would suggest a more professional tool like MATLAB (or one of its open-source clones like Octave or Scilab), R or Python with the standard scientific packages. Performance should be much better if you use them properly (vectorized calculations are much quicker than loops in all of the tools). And it is much easier to figure out:

Code: [Select]

tolerance = measured_value * gain_error + temperature * gain_tempco * measured_value + offset_error + temperature * offset_tempco
than:

Code: [Select]

=$A3 * Constants!$C$44 + $X3 * Constants!$C$46 * $A3 + Constants!$C$45 + $X3 * Constants!$C$47

Plotting a million random points takes a few of seconds with R or SciPy/Matplotlib on Python 3. Pretty much instantly in Octave.

[bitseeker]

Thanks, alm. I'll check out Octave and Scilab.

[CatalinaWOW]

I've started on Octave.  I've had some experience with Matlab and find that it takes a different mindset in formulating problems.  No problem once you get it, but it is frustrating until that clicks.  Just like any large package.

I actively resisted using Matlab before I retired, simply because I wanted my work accessible to the broadest population, and my corporation did not fund a Matlab license for every engineer, let alone technicians and managers of technical work.  But spreadsheets were on every machine, even secretaries and other totally non math oriented positions.  They also didn't allow freeware such as Octave. 

Situations like this are one of the reasons that Excel compatibles survive in the face of more powerful tools.  You will reach a more sophisticated audience using Matlab and similar, but a much broader audience with Excel.  (Also note that someone proficient in Excel can make equations much more readable than your example above, not as good as the Matlab version, but significantly better.  And comments are quite possible in both, a great aid when the author can be bothered to include them.)

[cellularmitosis]

It looks like feedback.loop also uses Octave.

https://youtu.be/jcCC867gKXE?t=26m47s

[alm]

Yes, it is possible to do better with Excel. But I would argue what I posted is fairly typical for a large fraction of the Excel spreadsheets I have seen. I admit that I have not been in a company with people with advanced Excel skills that spend all day in it, like you might have in some financial firms. But I have worked with plenty of scientists and engineers, and what I posted was fairly typical. Actually the worksheet named "Constants" might have been Sheet2 or the constants might have been shoved somewhere below the data. Excel just invites this kind of code by hiding away features like named ranges. Even a lazy scientist programming in MATLAB would at least have named the temperature t and the voltage v. I am not sure if I have ever seen someone else use the Excel comment feature except for collaboration.

I took a course which used Excel for some moderately complicated equations based on a dozen or so of input parameters. After looking at a student's Excel worksheet and seeing the results were all messed up, the instructor proclaimed that it is impossible to debug an Excel worksheet. I proceeded to find the bugs in their worksheet, but it was definitely much harder than debugging a MATLAB program that did the same.

If you have zero programming experience, then MATLAB/Octave/Scilab might be a fairly steep learning curve. But if you have any experience with procedural programming, you should be able to get started quickly. Learning how to (mostly) get around loops through clever programming or the intricacies of indexing takes a bit longer. Many solutions you find for MATLAB will be applicable to Octave or Scilab with some minor modifications.

For me Excel (or other spreadsheets) is fine for simple problems with at most a dozen columns, a few hundred rows and no complicated math. But as soon as it becomes more complicated, I move it to a proper programming language. What if you want to do the same analysis for a hundred CSV files (common case for me). Sure, you could write some VBA, but at that point you have pretty much lost all advantages Excel had about being less complicated.

Just to be clear, MATLAB/Octave/Scilab are not my favorite tools either. I think the language is pretty ugly. But I recognize them as professional tools and have written a fair amount of code for them (mostly MATLAB). Excel mostly strikes me as a toy, despite said toy being used by millions of professionals.

[Conrad Hoffman]

I have to agree with the accessibility of Excel, and I use it a lot. It just doesn't take much effort to get things labeled, colored and commented. Heck, comments help me remember what I did. You can also extend things greatly by enabling the advanced analysis features and adding some VBA code. That said, I usually do a first tab with background and instructions and half the people I give it to respond with, "OK, I see the text, but where's the spreadsheet?"

[alm]

The question came up because of the performance of Excel with a large number of rows. Do you have a way to get reasonable performance with 1M samples? In my admittedly old Excel 2007 it refuses to plot more than 32k points (making me do the decimation, instead of doing it itself like every other math package). And scrolling past this graph with 32k points (not many in my book) takes a couple of seconds per redraw cycle. While Octave was able to plot 1M points (decimated, obviously) in a fraction of a second, never mind moving the window.

How do you even decimate data without aliasing in Excel? Something like the MATLAB/Octave decimate function that low-pass filters the signal before resampling.

[Conrad Hoffman]

Don't know what the limits are, but I just created a column of 70,000 data points, summed them and graphed them, with no trouble at all. Took just over zero seconds. Running Windows 10 and Office 2016 on a 4.2 GHz i7-7700. No doubt one smarter than me could write some VBA code to filter and decimate.

edit- OK, I put some Excel data smoothing filters up in a separate thread. They seem to work well on voltage history.

[kj7e]

Package arrived,  Ill setup and begin some tests this evening.

[kj7e]

cellularmitosis, any specific tests you would like me to look at?  I have a 2 hour warm up drift plot of the Geller ref and a 10 hour plot.  I am able to confirm my 34465A as sent from Keysight consistently reads about 5 to 8 ppm low from various references, including the Geller.  Ill work on posting the data I have so far today.

Edit, doing a noise comparison, LM399 based PDVS1 vs AD587LQ Geller vs LT1021 Doug Malone vs LTZ1000A KX Ref.  Photos and data forthcoming.

[cellularmitosis]

cellularmitosis, any specific tests you would like me to look at?  I have a 2 hour warm up drift plot of the Geller ref and a 10 hour plot.  I am able to confirm my 34465A as sent from Keysight consistently reads about 5 to 8 ppm low from various references, including the Geller.  Ill work on posting the data I have so far today.

Edit, doing a noise comparison, LM399 based PDVS1 vs AD587LQ Geller vs LT1021 Doug Malone vs LTZ1000A KX Ref.  Photos and data forthcoming.

Sounds great!  Yeah, I would have just suggested whatever tests you are familiar with, voltage, tempco, noise, hysteresis.  The specific tests aren't critical in this round -- this is more about giving people a real object to engage with, and to start grappling with solving all of the challenges which come with it ("how do I record and plot data?  How stable is my lab temp?  What are my EMI issues?").  I know that for me, just mailing off this box has kicked into motion several projects which I've been mentally toying with for years!

[TiN]

Who is the last person if the food chain?
Perhaps somebody can list the loop members and sequence, for us outsiders? 

[cellularmitosis]

After kj7e comes bitseeker and then RandallMcRee, which completes the original route.

However, we've had a few more jump on the bandwagon!  nikonoid, martinr33, and orin.

Finally, I've also recruited someone (Vacuuminded) from the volt-nuts mailing list!  He should be introducing himself in the thread soon

[bitseeker]

More members. Yay! Welcome aboard, everyone.

[orin]

Should certainly be interesting.

Since the recent hot spell (max of 29 deg C in the lab), my 731B has been reading a little low on the 34461A, so tonight I fired up both the Fluke 8845A and my SVR-T.  The results were surprising.

Using 100PLC and stats to average 100 readings:

	34461A	8845A
731B	9.999967	9.999980
SVR-T	9.999969	9.999981

The last time I compared them seriously was back in 2013:

https://www.eevblog.com/forum/testgear/geller-labs-svr-and-svr-t/msg282582/#msg282582

If anything, they've drifted closer!  The 731B has been powered almost continuously (we've had some pretty long power outages) and the SVR-T unpowered.  I certainly didn't expect them to be so close.

The 61A is two years since its last cal and the 45 is a year old.

[Conrad Hoffman]

Cool! I've found that power cycling the 731s is less of a problem than one might think. In fact, I don't even have batteries in mine anymore, just a large capacitor and parallel zener diode to emulate a battery's filtering and voltage regulating action. They won't work correctly without that, or a battery. I do the same thing on my 845 null detector.

[kj7e]

My Excel skills are lacking, I can send someone my raw data if they could overlay the temp and voltage plots.

There really was not much of a warm up needed for the Geller reference, 2 hour plot;


10 hour plot, A/C was on but the room held steady within 1dec C.  You can still see the influence of the A/C cycling.  I cant explain the two spikes, possibly overhead lights being switched on or off;


Temp over the same 10 hour time frame;


Ill have the noise comparison of the various references later.  I would like to also do a tempco comparison if time permits.

[bitseeker]

I would like to also do a tempco comparison if time permits.

No need to rush on my account. Feel free to do the tempco, too.

[Conrad Hoffman]

Sent a PM to get the 10 hour.
Thanks,
Conrad

[kj7e]

Sent a PM to get the 10 hour.
Thanks,
Conrad

Email sent.

[Vgkid]

Does anyone have/want to do a low ohms test.

[nikonoid]

I have two micro ohm meters (510a,Simpson). It might be a good test.


Sent from my iPhone using Tapatalk