[Diy_efi] NTK wideband sensor

[Garfield Willis]

On Thu, 7 Nov 2002 09:14:34 +0800, Bernd Felsche
<bernie at innovative.iinet.net.au> wrote:

>> The above should read "at each AFR *for* a statistically significant
>> sample".
>
>Oh... how long does it take to conduct an infinite number of tests
>on a statistically-significant number of samples?

Silly and pedantic as usual. What's the resolution of AFR measurements
currently stateODart? Hmm? Oh about 0.1AFR, so over a range of from 10
to 18 AFR, how many points minimum in that spread over "each AFR"?;
that's about 80 points [we want 0.1AFR resolution, and so there are 10
such points in one AFR, and our measurement range/spread is over 8 AFR
(10 to 18)]. In our full-sweep gas calibration data, we take 1000
points. Well over 10X the minimum required for 0.1AFR resolution in the
cal. A goodly number of points on the curve, but certainly not an
"infinite" number. Plus, with that many points, we don't have to
interpolate. And for that many data points, with decent, only slightly
automated gas equipment, it doesn't take all that long to test each
sensor. We do it to every one we ship. Nothing "infinite" involved.

>How did you determine what was a statistically-significant sample
>size?

Same way ANY data collection process does. Do YOU know how it's
done/determined, by any chance? (altho you clearly meant it to sound
pedantically rhetorical, just like your previous question, it's clear
you also don't know how a 'statistically significant/adequate' sample
size would be determined when taking data samples in general. Do a
search of the web and you'll see I didn't pick that phrase outta thin
air. Standard nomenclature for determining a distribution from measured
data. If the mfg. doesn't give it to you, you have NO other recourse but
to do due diligence and measure it yourself.

Since we can't know the # of sensors being mfg'd over say a year's time,
determine a desired confidence level, and then cull out the required
samples at random, we have to use some estimates. Let's guess that since
the sensor was endOlife'd in OEM production usage long ago, and so the
current mfg. volume must supply for repair parts and a probly much
larger aftermarket volume in proformance gear, we can pick some number
like 10K sensors per year. OK, so even IF we could pick samples from
these randomly (which we can't because you can't avoid the fact that
they're clearly made and supplied in batches; witness the current
shortage), to reach a decent 95% confidence level in our distribution,
we'd need to test about 300 of them. That gives you an idea of what
level of testing is involved in a "statistically significant sample". A
number that's quite a bit bigger than one or two, or even a dozen.

The point of all this shouldn't be to sneer at the process, but to
realize that trying to determine a calibration curve from just a FEW
sensors isn't going to get it. That was my point. You tossed out a
couple of straw-man rhetorical questions, seeking to poo-poo these
issues. To what end? to suggest that it's perfectly OK to throw a dart
at the wall and call that a calibration? T'would seem so. But like I
said, that's just silly. Nobody is served by such pedantry.

Gar


_______________________________________________
Diy_efi mailing list
Diy_efi at diy-efi.org
http://www.diy-efi.org/mailman/listinfo/diy_efi