NBS 430 and alternative methods

[ManateeMafia]

I have been working on a project that uses the information contained in NBS Technical Note 430 aka NBS 430. http://www.dataproof.com/NBS430.PDF
It is a great document with examples on how to calculate the emf value of standard cells. This document can also be used for the monitoring of other voltage references like the Fluke 732A/B, Datron 491x, etc...

While I have made progress with the document, I have also come across reference to other methods of doing the same thing. The Data Proof website mentions these methods in the Voltref web pages. It is on page 2 at the following link...
http://www.dataproof.com/VoltRef.PDF

The quote is below:
Choice of Designs
VoltRef provides a choice of three test designs: 1) statistically balance design recommended by NIST and used by the majority of western national laboratories, 2) favored cell used by some European national labs, and 3) ring design popular in a few Asian countries. (Available only with VoltRef for Windows.)

I would like to know if anyone can point me to the last two methods used in Europe and Asia. If possible, I would like to see how well the different methods compare.

Thanks for any help.

[dacman]

The other methods can still use the formulas in NBS TN 430.  Favored Cell reduces the number of measurements by only performing the measurements required for one (favored) cell.  The Ring System allows a system with more than five cells to use the five cell formulas and methods.  (The five cell system is efficient and produces the lowest uncertainties compared to the three and four cell systems, and is balanced, unlike the six cell system.)

In NBS TN 430, an intercomparison of three cells takes six measurements and any one cell gets four measurements.  For four cells, there are twelve total measurements and any one cell gets tested six times.  For a five cell system, there are ten total measurements and any one cell gets tested four times.  I don't like the six cell system, because it is not (completely) balanced.

A favored cell system reduces the amount of measurements by only taking the measurements needed to calculate the pooled value of the favored cell (and the formulas in NBS TN 430 can still be used).  For a five cell system, only four measurements are required (which are the measurements that the favored cell is used in).  Then the pooled value and pooled uncertainty for the favored cell can be calculated.  Tomorrow you could pick another favorite.  The other cells could still be used to their (predicted) values and uncertainties.  (There hasn't been enough measurements for the other units to have a pooled value.  For one, for the five cell system, each one would have only been used on either the A or B line and only have one measurement, meaning the system was not balanced for them.  Also, more measurements would reduce short term drift and noise.)

A ring system is for when there are more than five cells in a system, otherwise it is the same as the default system.  If the cells were placed in a ring, any one cell is compared to the two on its left and two on its right.  Then the five cell formulas could be used.

The most basic form of pooled uncertainty is if the uncertainties of the (predicted) values were all the same, then the (average) uncertainty could be divided by the square root of the number of cells involved to arrive at the pooled uncertainty.

[ManateeMafia]

dacman,

Thanks for the response. I was thinking about the favored cell approach to use with an external standard. I wonder if this is the preferred method for a cal lab to transfer the volt from their array to a customer or at least to monitor it over several days after using a direct series opposition measurement using a single standard?

The lab would use the standard method for their own references but use the favored cell approach for everything else. I can always write the routines to do this and check the results against what I am using.

Also, I wrote the code to use the same pattern used by the 3,4, and 5 cell example. It extends the same pattern as large as needed. It gets to be quite large for a 10 cell array but I wasn't sure if the approach was an issue other than the time needed to complete the scan.

Currently, I have two smaller groups of 4 cells and 5 cells. I then run them all together for fun.

[dacman]

If time is of value (especially if there is more work than there is time to do it) I can see using the favored cell method.  (I have to transfer results to another system, check plots, and make sure things are in control.  However much a metrologist wants to check all the cells, management might not allow it unless it is required.) The other calibrated cells could still have trends based on the reported values.

There are labs that calibrate one cell and use other cells for check standard measurements to see if their lone calibrated cell is ok.  Pooled uncertainty would not apply in this case.  (VoltRef can be used even in this case, although it requires at least three cells to run.  Only one needs to be calibrated.)

I probably gave too much of my opinion in my previous post.  In VoltRef, even the six cell system is balanced because VoltRef uses about 20 measurements, 10 on each channel, then combines the results, for one term in NBS 430, meaning each term is balanced.  If you are paying high priced technicians to do the work, and this is done regularly, VoltRef will pay for itself in short order.

[dacman]

One use for the cell with a pooled value would be to take it out of the system and use it to calibrate a Keysight 3458A or Fluke 5720A.  The pooled uncertainty could be used to determine if the cell was accurate enough for this.  The other cells would not need a pooled value.  (And a well behaved Fluke 732A or 732B would not need a pooled value either.)

To calibrate an unknown cell with a bank of cells, the pooled values of the cells would not be used.  The comparison values would use the calibrated (or predicted, if applicable) values, and the uncertainty of the measurement would be based on the uncertainties (or uncertainties of the predicted values) of the calibrated cells, which could be pooled together.  To calculate a measurement uncertainty for the unknown cell, standard deviations of the measurements would need to be factored in, which could be obtained by taking multiple measurements for each term.

One reason to obtain pooled values for the cells (for all of them), is if the cells are being trended and predicted (from their calibrated values), then (for an accredited lab) the cells are supposed to be checked to see if they are actually trending within the uncertainty of the prediction.

[ManateeMafia]

dacman,

Thanks again for your help. Initially, paying for the VoltRef software was not considered an option since I wanted the experience to control multiple items and do the number crunching myself using the examples. The cost of the software certainly would have been much cheaper in terms of labor and compiler renewal costs. The more I write, the more I realize there is a lot more information needed to get better use of the potential data.

I will be adding the pooled value option in the near future and will test it with another member with a similar setup. I haven't seen anyone mention any automated data collection from references even though I know a few members are slowly building up a collection of LTZ boards and standards from eBay.