Re: Weibull Distribution but with a varied cycle time

The classic Weibull analysis is concerned with time-to-failure data or
stress-to-failure data. Or cycles to failure, for that matter, since
that is a form of stress to failure. The spring data is an example of
stress-to-failure data. I believe you are anticipating time-to-
failure data. In a Wiebull analysis both are treated the same way.

Getting appropriate and sufficient time-to-failure data can require a
lot of time because you need to experience a fairly large number of
failures in order to do a reasonable analysis.

So the concept here is to put a device or system into service and
observe how long it runs before it fails. This is often done with the
device or system under an abnormally large load in order to accelerate
failures... in which case you get accelerated failure rates. Of
course, if you do that, your data is not consistent with failure rates
under normal loads.

Also, as the article mentions at the end, there is the matter of time-
to-failure data combined with tests that ran on and on and the device
or system hasn't failed yet. That's called censored data. This is a
common outcome. Don't neglect the censored data. You can't just
ignore it. The censored data (ran on and on and hasn't failed yet)
has valuable information. If you just ignore this information that
will bias your analysis and cause it to be misleading.

I have a couple of suggestions. First, Quality Digest is not a very
upscale magazine. Articles in there are usually written by someone who
is selling products or services. They usually tell you just enough to
make you interested in hopes you will buy something from them.
Second, the last way I'd go about a Weibull analysis is to use Excel.
There is a wonderful thing called "Weibull graph paper" that gets
around all of the opportunities to make mistakes with Excel. Rank the
time-to-failure data... get out a pencil... sharpen it... and plot the
data on Weibull paper... then interpret it. Ahhh!! There's the matter
of interpreting what it is teaching us. The author of that paper came
to the edge of this, but didn't drill into it. He wasted a lot of
paper on the mechanics of using Excel, but didn't explain the
remarkable things we can learn about our device or system by simply
plotting time-to-failure data (or stress-to-failure data) on a piece
of Weibull paper. Weibull graph paper is designed to lead us directly
to these insights. A homemade sort of graphic via Excel doesn't do
that.

If by some chance you decide to collect some time-to-failure data and
want an analysis the easy way, let me know. Post here and send your
post to me as "Send to Author" or whatever that button is. It will
take about five minutes to plot the data and a little more time to get
the plot back to you with an interpretation. No charge for this sort
of thing, of course. Before you actually get the data it might be a
good idea for you to document a little more about the device, system,
or object you are studying. Let's be sure that the thing you are about
to do can produce valid and useful data.

Time-to-failure data and all sorts of variations on it are a whole
subject unto themselves. Especially if some of the data is going to
be censored.

OMU (I'm an engineer and a statistican, just fyi)



On Jan 27, 1:59 pm, "googlinggoog...@xxxxxxxxxxx"
<googlinggoog...@xxxxxxxxxxx> wrote:
Hi,

Im no expert at statistics as the following question might prove but
i'll try to explain what I wish to achieve.

Basically for work i would like to perform some reliability tests on
an electronic unit, the tests will include powering it on and ensuring
that the outputs are as expected. I am going to automate this part of
the test.

But my question is, having read this article (http://www.qualitydigest.com/jan99/html/body_weibull.html) to reinforce my
existing knowledge it uses the explain of a spring, this only really
has a set duration of operation, ie - when the box is opened it
expands, and prehaps fails.

Now my problems lies here, my unit could fail anywhere turn on to
24hrs into the test, so thats partly my question will my weibull
function hold up for this? do I need to test for the average period of
usage? becuase testing based on cycles seems stupid to me for my
problem becuase i could turn it on and off 100's of times and it might
fail, alternativly i could turn it on once and leave it for hours/days
and it might fail, but then it would only be one cycle...

So really I suppose im after some pointers. Obviously i need to define
a period to test for and that it can operate for without problems I
think.

I know my questions somewhat vague, but thats due to not truely
understanding the best way to go about this.

Regards

David

.

Loading