Using Curve Fits and Predictions in place of good old Measurements

[Rmack]

Hello,
I’m hoping for some input on an ongoing practice at my lab. Over the years the engineers have developed a program that does a least means squares curve fit for sensor data. This has evolved into becoming the tolerance for calibrations. For instance, the prediction error of a curve fit with 3 coefficients would have to be 0.1% or better for the UUT to be considered good. It doesn’t matter what the readings were, pass or fail relies solely on the prediction error.
Being a traditional kind of guy this makes me uneasy. I like it when each reading has clear plus or minus tolerances. Making sure that each reading is what it should be.
I can see using the prediction error as a usability spec for our internal customers but it doesn’t seem like it should be the pass/fail criteria for a calibration.
The thought of having to calculate the uncertainty of measurements done using this method gives me a headache.
So please, I’d appreciate some feedback on this. Am I just behind the times and need to get with the modern age or are we up to something that’s not quite right? We’re working towards 17025 accreditation and this has become a big concern for me.
Thanks,

[Stijloor]

A Quick Bump!

Can someone help?

Thank you very much!!

[Steve Prevette]

Not my primary area of expertise, but I'll take a stab at this. The advantage of using what you describe is that it integrates the errors across all of the readings taken and comes up with an answer. The problem with only looking at individual errors is that a few of the points being read may not be far enough off to fail calibration, but the overall effect is that it should.

If you have the time, I'd sit down with whatever spreadsheet calculates this and put in a few trial "errors" and see what it takes for the routine to flag a failure. Then you can get a feel for how large an individual error is needed, and also the effect of being near the center of the curve versus "out on the edges".

[fast_hedgehog]

This is fine...
You can imagine a simpler situation where you are calculating a first order least squares curve fit. Deviation from this line can give you nonlinearity. Its common for transducers to have accuracy specified as nonlinearity, and this is often the only way to determine "pass/fail" for a transducer.

There is no reason that you shouldnt do similar with higher order curve fits.