On Sunday, September 11, 2016 at 2:14:50 PM UTC-4, FromTheRafters wrote:
trader_4 pretended :
On Saturday, September 10, 2016 at 10:17:50 PM UTC-4, T wrote:
On 09/10/2016 12:41 PM, Ed Pawlowski wrote:
tolerance for MTBF

Hi Ed,

That is not what MBTF means. If the bulb had, say 1000 hour
MBTF, it would mean that you put 1000 bulbs in a test bench
and ran them for an hour. Only one failed. MBTF does not
tell you anything about the second and so forth hours.



This is just nuts. You can't get any meaningful MTBF by the method
you just described. What you showed was taking 1000 bulbs and testing
them for just an hour. That isn't predictive of MTBF over the life
of the bulbs. Following that method with say an ordinary incandescent,
you'd come away with the impression that they hardly every fail at all,
because incandescents rarely fail in the first hour and it has little
to do with how long they last in service. (In reality what you'd be
measuring is the infant mortality by your 1 hour method)


You find MTBF by testing the devices continuously over many hours,
finding out how many fail at 100 hours, 1000 hours, 10,000 hours
and then determining on average how many hours you get before failure.
Ed has it right, MTBF, properly calculated, is the average number of
hours that you get from an LED before it fails. In your bulb example,
you only tested one bulb to the failure point, in essence you have a
sample size of one.

"MBTF does not tell you anything about the second and so forth hours."

Of course it does. If we know that a bulb or an engine has an MTBF
of 20,000 hours, then we know that on average, that's how many hours
they go between failures. The device is very unlikely to fail at
two hours, or two hundred hours, but has a high failure rate at 20,000
hours. Are you telling us that MTBF only tells you about the first
hour?


(I did MBTF analysis for the military.)

That's scary. And if that's the case, why is it that every time you've
used the term here, you keep posting "MBTF", when it's actually MTBF?

I'm glad somebody noticed. The 'mean' time of a single failure is
exactly the time of that failure and is basically useless as a measure.


+1


I would think that they could test 100 items until maybe 25 of them
failed and get the mean time from that selection of failures.

That would be getting closer to the real number, but even that gives you
an MTBF number that is too low. That's because you're only counting
the 25 that failed and not counting the 75 that were still working OK.
You know the MTBF based on averaging the number of hours it took for
those 25 to fail is a lower number than what you would calculate based on
averaging what you'd get from all 100. Knowing something about
the physics of whatever the system is, eg that it would typically follow
a bathtub shape curve, you could use some statistical methods to generate
a better estimate of the real MTBF based on the results of just 25 failing
out of 100 and stopping the test at that point.