Deriving design precision specs. (µCurrent GOLD example)

[Christe4nM]

Hi Dave,

Lately I've been wondering how you determined the precision of your µCurrent design. It is more than just the µCurrent, but in general how do you get from component precision to circuit precision. I thought that it would make a great fundamentals Friday. Just in case it's not, could you still reply, 'cause then I might open a topic in the projects section.

Keep on doing the great work

Christean

[c4757p]

I agree, error analysis would make a great FF topic.

[robrenz]

+2

Including how to calculate your final accuracy including voltage specs of whatever you have µcurrent plugged into.

[Wytnucls]

The MetraHit 30M has a resolution of 100 pA on the 100 uA DC range, with an accuracy of 0.02% of reading + 0.002% of range. What kind of DMM does one have to plug in the uCurrent to achieve the same kind of accuracy?

[EEVblog]

That's a big can'o'worms potentially.
In this case I simply wanted "better" than my old one, so chose the next best precision value available (0.05% instead of 0.1%).
You can calculate "circuit precision" from the datasheet values in theory, but in practice I test to a tighter spec than what the datasheet says.
But then you have the whole issue of metrology and calibration, drift, aging, temperature etc.
It would be a video where it would be guaranteed I would:
a) do something different to "how it should be done" and hence the complaints
and
b) brush over many areas for the sake of simplicity leading to more complaints!

In a word, from a blog video point of view, I think it's pretty a pretty "messy" topic.
But I agree it's an interesting topic.

[c4757p]

Of course there'd be whining, but isn't there always? It's a complicated topic, but that's kind of the point to making a video about it!

[tszaboo]

but in general how do you get from component precision to circuit precision.

There is the "worst case" analysis, where you have to know how your circuit works, and add together the errors (with significance). There is the monte-carlo analysis, where you just run a simulation few thousand times to get a distribution, and then make confidence intervals for the accuracy. You can also test a few dozen finished product. Also, there are several methods to determine which parameter is significant, which is not.
This is definitely not beginners topic...