Success testing

[longlife]

Success testing to one life seems to me to be a very inefficient method for demonstrating a minimum reliability level. I am trying to understand when this method might be useful. Seems to me that with a little effort, a reliability engineer should almost always be able to propose a more efficient (time, samples, etc.) method for demonstrating reliability.

What do the rest of you think?

[Pantelis]

I am not sure about using the term efficient, I would rather use the term practical. Depending on what you are testing it may not be practical. In the cases however that it is practical, testing over the whole life of the product provides a much more informative and robust reliability estimations as it minimizes any assumptions (such as acceleration factors).

Now with regards to pure success testing (i.e. zero failure success plans), I would also argue that, if practical to test to one life, why not extend the testing past this point. Extending to a point where you begin to see failures becomes even more informative.

[longlife]

By inefficient, I was considering the resources expended. Success testing (no failures) to just one life requires the least amount of knowledge about the component being tested, the expected fail modes evaluated by the test, the failure behavior (life distribution) of the component etc. If the reliability engineer puts forth a little effort to understand the test, the component, the expected fail modes, and their distributional characteristics, then an accelerated test to failure can be developed. The developed test can provide the same reliability demonstration with fewer samples, less test time, fewer test set-ups, etc. I am having trouble conceptualizing a situation in which proposing a success based test to one life would be the appropriate option.

[Pantelis]

Well, first it is highly dependant on what you are testing. In the case of one-shot items it is an obvious choice, but it would also be a good choice in the cases of where life can be accumulated fairly quickly. Given this case (i.e. short/reasonable test duration) it would also be prudent to run the test longer (past one life) until failures are observed to better characterize the failure rate behavior. Furthermore there are also the cases where, at the system/subsystem level testing, acceleration may not be possible. In other words there are many reasons for going down this route.

Additionally one should not loose sight of the fact that a well designed test at use conditions (be it to one or more life) provides higher confidence in the results than a comparable accelerated test. In general and to achieve the same level of confidence an accelerated test would require a larger sample size.

[longlife]

I had not considered one-shot items. A very good point.

I do not follow for cases where life can be accumulated quickly. Rather, I agree with your suggestion to run the test longer until failures occur.

In cases where acceleration is not possible ( or too difficult ), I see another good point. This may also be a situation suitable for success testing to one life.

I don't follow the last paragraph. Even without accelerating a test, if the failure distributional characteristics are known, a given level of reliability can be demonstrated with fewer resources (samples, test time) than if nothing is known and the success to one life scenario is employed.

[Pantelis]

With regards to the first item I was referring to usage rate acceleration. In other words if a motor only works 30 minutes a day one can accumulate life fairly quickly.

I am afraid I don’t get your last one … what I was trying to say is that an accelerated test (using elevated stresses) will in general require a larger sample size than life test (tested at nominal) to obtain results at the same confidence level.

[longlife]

I posted a reply (quite lengthy) but the system lost it. During this session alone I have had to log in 3 times. When I posted my reply, the system claimed I was not logged in and lost my post.

I will try again on another day.

This is not the first time this has happened.

[longlife]

Pantelis,
I understand now what you meant by usage rate acceleration.

The last part goes to the heart of my original post.
...Even without accelerating a test, if the failure distributional characteristics are known, a given level of reliability can be demonstrated with fewer resources (samples, test time) than if nothing is known and the success to one life scenario is employed.

If we ignore any acceleration and only test assuming no failures allowed, the test plan based on the greater amount of relevant information will cost (time, samples, etc.) less. With no distributional knowledge, the binomial distribution can be used and some number of samples determined with the test run to one life. If the distributional characteristics (Weibull slope for example) are known, the test can be run to more than one life and fewer samples are needed to demonstrate the same reliability level at the same confidence. Just by adding some information, the test is made more efficient (less costly).

My original contention was that the success test to one life is almost never practical because a competent reliability engineer should be able to gather some relevant information that can be used to make the test less costly.

[longlife]

No other thoughts on this topic? I am a bit surprised.

[Pantelis]

I was out for a while... and I guess nobody else sounded in...

Anyways, I agree that if you did a life test with no acceleration you would gain much more useful info than a zero failure test plan, and with less samples. The issue is that all things have their place and depending on what you are doing you may find that one approach is preferable over the other, as I pointed out in some examples. Furthermore I will agree with you that one should not just apply this universally without regards to many other factors, and to take this a step further and say that more often than not it is not be the best choice. In other words all have their pros and cons and one should consider these when setting up a test plan. What I have an issue with is blanket statements such as never practical… or the opposite …

At the end of the day, and regardless what testing method you use you will have to choose two of the following three: small sample size, short test time and “good” results… You’ll never get all three!

Pantelis

[longlife]

Thanks my friend. I appreciate the thoughts. I too try to never say never. In this case though, I find too many people claiming to be reliability engineers but really aren't. The use this success testing to one life as the only test plan and consider themselves experts. The cost to companies following this misguided approach is appalling.

Catch you on another subject.
Mark

[Oleg_I]

Mark,
You have answered your question yourself. "If the distributional characteristics (Weibull slope for example) are known, the test can be run to more than one life and fewer samples are needed to demonstrate the same reliability level at the same confidence. Just by adding some information, the test is made more efficient (less costly)."
It is used for product's life warranty in many industries (for example aircraft engines).
There is a tool for this case: http://win.mail.ru/cgi-bin/getattach?file=LCsimulator.ppt&id=12772969900000000251;0;1&mode=attachment&channel=&notype