MTBF from ALT results

[roberta]

Dear all,

I have a problem in planning my test. The product is a module for a PC card. From a kind of HALT (we tested just one product...) the unit reached 105°C, and after that we stopped the test because the SIM or the antenna used to monitor the sample start to melt or have other errors. My problem is the following:
I have 20 units to test, to be divided in 2 stress levels (10 in each one).If I make an ALT suppose at 90°C and 95°C as stress levels, and I don't have any failure during the test time (that unfortunately cannot be more than one month) or very few (e.g. only two failure in each level and with high TTF), what's happens? In the sense, how can I calculate from these kind of results (I use ALTA6 basic version) the MTBF of my product?

Thanks a lot for your help, I really need it!!

Greetings
Roberta

[Pantelis]

Roberta,

The basic purpose of a quantitative ALT is to determine the effect of stresses on life, and from that infer a life characteristic at use conditions (different stress). To be able to quantify the effect of the stress on life at each condition one must observe some failures (end of life events). Unfortunately with out observing that, no model can be built. It is possible to built models with as little as one failure, however the accuracy of any inferences made with limited data is always questionable.


Now what else can you do… You can assume any of the parameters you can not compute due to limited or nonexistent data (such as an acceleration factor, etc.) and use these assumptions to get your result. The drawback to this is that the validity of your answer is highly dependent on the validity (correctness) of your assumptions.

Hope this helps.

P.S. Other options also exist, such as degradation type of analyses, assuming that you have something you can measure and model under these stresses. Since this was not mentioned, I am assuming that this is not an option.

[roberta]

Degradation could be an option if I could monitor the bit-rate or the loss of data (for example). The fact that I'm testing wireless product, so monitored with an antenna and with the SIM card inserted, makes the test more difficult. For example, if at T my product fails, because the SIM fails....what can I say? Must T be considered as a DL?

For what I saw....I don't expect sufficient failures to build a model. Anyway, testing at high T, maybe we have enough failures but the extrapolation is too high!!

This is the typical case when you have to face with situation where the egg and the chicken are both wanted!! :(

I really appreciate your help and support, because I'm quite new in this job....but I love it so much and I would like to succeed!!

Many thanks for your time

Roberta

[Pantelis]

Roberta,

I am not sure I understand if there is a question in there ... or your making a comment on my prior posting.

[roberta]

yes, is actually both!!
In the sense, I understood how generally works the degradation analysis, but I don't know how can I perform it on wireless products. Mine was just an example (bit-rate, loss of data, etc..), but I don't know if it's doable in practice. I have the problem that the customer asked for an MTBF calculation, but I have only 20 units (10 divided in two stress levels) to be tested, and maximum 8 weeks of testing time available. I don't know what I can conclude in case of very few failures (or no failures at all).

Thanks a lot

Roberta

[Pantelis]

With respect to the specific issue that you are mentioning, I am afraid I am not an expert on wireless products, and can only answer you in generalities. See below:

First degradation analysis involves the measurement and extrapolation of degradation or performance data that can be directly related to the presumed failure (mode) of the product in question. See http://www.weibull.com/LifeDataWeb/degradation_analysis.htm (http://www.weibull.com/LifeDataWeb/degradation_analysis.htm)
See also http://www.weibull.com/AccelTestWeb/degradation_analysis.htm (http://www.weibull.com/AccelTestWeb/degradation_analysis.htm). I am not sure if you can or cannot do it based on what you have.

Now what else can you do with what you have:

First, resolving an accelerated test with no assumptions, but just from data without any failures is not possible. What you can do however is make some assumptions and use other methods.

Let me at first address what one can do in the case of regular life test (not accelerated).


Option 1: assume an exponential distribution (constant failure rate) and infer the reliability (or your metric of interest, in this case MTTF) based on the non failed data. This is fairly easy to do (fit an exponential distribution to the suspension data). The problem with this is the constant failure rate assumption. While this is a valid assumption for the period of time you have tested the product, extrapolating beyond that may or may not be realistic since wear out mechanisms may surface at a later time. More than likely what you want to infer is beyond the test time, thus the problem.

Option 2: Is to assume some failure rate behavior. You assume one because with no failure data, a failure rate behavior cannot be inferred from the data. You do that using models such as the one parameter Weibull (or if you want to make it more complicated using a Bayesian Weibull model). See http://www.weibull.com/LifeDataWeb/the_weibull_distribution.htm (http://www.weibull.com/LifeDataWeb/the_weibull_distribution.htm) and http://www.weibull.com/LifeDataWeb/weibull-bayesian_analysis.htm (http://www.weibull.com/LifeDataWeb/weibull-bayesian_analysis.htm). The issue with this is that your analysis is again as valid as your assumption with regards to the failure rate behavior (i.e. beta value in the one parameter case, or the distribution of beta in the Bayesian case).
Now both of the above cases deal only with one condition (single stress) and not an accelerated test. In the case of an accelerated test, with no failures, no inference can be made since from the data you have no idea what the effect of the stress is on life. Now lets say you know (assume) the acceleration factor (X) between condition 1 and 2, and you tested at 2 then if you analyze the results at this level as per option 1 or 2, your results at 2 (higher stress) can be multiplied by the Acceleration Factor X, i.e. MTTF(1)=MTTF(2)*X, to obtain that metric at the lower condition 1. The issue with this again is the possible errors mentioned earlier, plus any error in your acceleration factor assumption.
I know I got long winded on this, but I hope it answers your questions.

[roberta]

Your answer is clear. My problem is that we have to do the test at two levels (since around 100°C the SIM gives error we have to stay below this value), suppose 85°C and 95°C. I tried to make just some possible example in ALTA/Weibull++ with some failure times (within my max test time that is 720h for each test at each stress levels). Since I don't expect early failures I put high failure times and many suspensions at 720h. In this cases, suppose 4 failures and 6 suspended the R(t) or the MTBF are very low.

If you want I can send you by email an example that can be better understood.

Thanks a lot for your help.

Roberta

[Pantelis]

I understand what you are saying .. what I dont know is what you are expecting me to answer. There is no magic, if you cant prove the reliability given your constraints in time, sample sizes and stresses then you cannot. Its as simple as that. There is no magic.

Now on the second line you said something about the MTTF being low. Well you computed that using assumptions. If the assumptions are right then this is the value, regardless if its low or not, or whether you like it or not. If the assumptions are wrong (your failure times that you entered) then the answer is wrong.

[roberta]

I understood completely what you mean. Is just that I'm working with very few information about these products, and in this cases I'm ,ot really sure that doing an ALT is the right choice. But, since prediction is even worst, the choice has been to do an ALT!!
We will see, I hope in reults differents from what i expect!!

Many thanks for your help, very useful!

Cheers
Roberta

[Pantelis]

Good luck Roberta.

I am now closing this thread.