Explaining the concept of step-stress testing to a Customer

[thiravia]

I recently developed a step-stress test plan as a part of the reliability demonstration to a customer. This consisted of single step-stress profile with approximately equal steps. However the customer objects to using a single-stress profile as he thinks it would only give a measure of the life distribution (variation within units) and cannot estimate the parameters of the relationship..since they are subject to the same step-profile..
I tried to explain to him by..
1) Pointing out that it is true only in constant stress and in a step-stress you would have different levels of cumulative damage (since they fail at different times) thus enabling estimation of acceleration model parameters.

2) Showing how the software can estimate results for such profiles. (using hypothetical failure times for my test plan)

3) Referring to several published papers where a single step-stress has been used.

4) Note from Reliasoft technical support where they stated that “while two step-stress profiles is recommended and can provide additional information one profile should still provide valid results”.
None of this seems us convince the customer!! Anyone with similar experiences or suggestions?
Although I may be able to come to some kind of a compromise in this case by adopting a two step-stress profile. I still need to reconcile the differences as there are other tests where having two step stress profiles may be prohibitively expensive due to test equipment.
Amar

[Pantelis]

Amar,

What you’re describing here is basically two issues. To put it simply the first is “is it solvable?” and the second is “is it good/acceptable?”. The fact that it is solvable, as you correctly pointed out, does not necessarily make it good or acceptable I don’t know any more details than what you provided in your post, however the more “stress profiles” you have the “better” the results. As a matter of fact I am not sure if you computed confidence intervals on your answers, from the analysis you did, but I would venture to state that they are huge. In other words even though the analysis is doable, the uncertainty that is related with any results you get is great. Using more profiles reduces that. My guess is that, this is what your customer is objecting to, and rightly so.

To put the above into perspective you can also solve a 2-parameter Weibull MLE with a single failure point (and say one or two suspensions). The issue here is again similar, that is, even though solvable the uncertainty in the modeling may not be acceptable.

Hope this helps.

[thiravia]

Thanks. You make a good point.

The issue of confidence intervals was known and deliberated during the planning, and with the trials that we did two profiles did not seem to improve it much with the samples we had. so we decided to look at B10 life as a conservative measure in addition to the median life. (not sure if anyone has done this..)

But the customer's objection in this case is however that it is not solvable. like I mentioned in my orginal post their argument is that since all units are subjected to the same profile , the different failures can only show the variability and not life stress relationship.

[Pantelis]

Well,
With regards to the solvable issue then it’s an academic question. The fact of the matter is that it is. Point them out to the math and they can follow it from there: A good starting point is:
http://www.weibull.com/AccelTestWeb/time-varying_stress_models_chap_.htm (http://www.weibull.com/AccelTestWeb/time-varying_stress_models_chap_.htm)
This is not a discussion that has a gray area. It either is or it isn’t. Given a mathematical methodology spelled out that does it, I don’t see where the argument would be on this point.

[Pantelis]

Let me add a secondary comment to the above.


Basically the only argument is mathematical... in other words here are the mathematical steps of how you do it (above link). If there are no errors in the formulation then you can do it. If they want to argue the other side, then they need to point out where the mathematical error in the formulation is.


On that point there is no error, and I would be more than happy to get involved if they say there is one.

[thiravia]

Thanks again for your thoughts. Yes , this has turned out to be an academic excercise at this point.

But the customer says that we dont understand the math ! :)

I think the disconnect is that they look at this as a graphical method method and struggle with the concept of constructing two points from one profile. I intend to educate them about the MLE method and perhaps write out the differential equations that when solved would give the parameters.

We may seek your help as a consultant if we run into some obstacles.

[Pantelis]

Good luck :-)

Let me know how it turns out.