SRC Forum - Message Replies
Forum: Reliability & Maintainability Questions and Answers
Topic: Reliability & Maintainability Questions and Answers
Topic Posted by: Reliability & Maintainability Forum
Organization: System Reliability Center
Date Posted: Mon Aug 31 12:47:36 US/Eastern 1998
Posted by: Johnson Wood
Date posted: Tue Aug 14 3:55:01 US/Eastern 2001
The circuit boards burn-in datas are recorded per month, judy as follows:
Date Sample Size #Failed
2001-2 1494 34
2001-4 3494 144
2001-5 25902 702
2001-6 20082 682
Which method of Weibull is suitable for this type of data?
Subject: Weibull Data
Reply Posted by: Joe Dzekevich
Date Posted: Tue Aug 14 16:28:06 US/Eastern 2001
No method of Weibull is suited for the data. Weibull, like Normal, is a continuous distribution. Your data is discrete, # good, # bad. For a continuous distribution like Weibull, you need a time to failure for each board or run time without failure per board (censored data).
What are you trying to find out?
Reply Posted by: Johnson Wood
Date Posted: Tue Aug 14 21:54:16 US/Eastern 2001
All of our circuit boards will experience a 55 C, 24-hours burn-in just after assembly and production.In the process of burn-in, some faults occur. I want to use weibull technology to predict the infant mortality reliability of life of the boards base on these datas. I bought the New Weibull Handbook from RAC, and I have the SuperSMITH Weibull tool also. Any advice will be welcome. Thanks.
Subject: Infant Mortality / Weibull
Reply Posted by: Joe Dzekevich
Date Posted: Thu Aug 16 9:33:47 US/Eastern 2001
OK...I see now. You want to build-up a distribution and/or histogram of the infant mortality distribution. Break your 24 hours up into multiple time periods, let us say three 8 hour shifts, and collect your data. Write down how many boards are running and how many are up at the end of each 8 hour period. Analyze the data as "inspection" or "interval" data. Treat a runing board as "censored" and a failed board as "failed" and use the time at the end of each inspection period (8, 16, 24 hours). You may also wish to do a special extended test. Run the burn-in for multiple days, taking data on each 8 hour shift. I can de dangerous to extrapolate the data too far into the future, so a multiple day run may help you there. Do this for eacl burn-in lot, and you will construct multiple Weibull distributions.
Subject: Weibull reliability estimates from burn-in data
Reply Posted by: Larry George
Organization: Problem Solving Tools
Date Posted: Tue Aug 21 15:12:39 US/Eastern 2001
The least squares estimates of the Weibull parameters are:
alpha = 9512 hours and beta 0.577 with MTBF of 15,100 hours.
The max likelihood estimates of the Weibull parameters are:
alpha = 9330 hours and beta = 0.582 with MTBF of 14,600 hours. My parameter notation is defined by the Weibull reliability formula R(t) = exp(-(t/alpha)^beta). Contact me for information about the estimators.
The agreement of the lse and mle estimates lends some credibility to them, but extrapolating from 3% failing in 24 hours to ~15,000 hour MTBF stretches credibility, even for MTBF predictions. The ~15,000 hour MTBF is on the low end of PC board MTBF predictions.
Do follow the recommendations to record failures at various times during burn-in. Also, extend burn-in for samples to get longer run times for some. This is known as "on-going reliability" test by some, "rolling reliability" by others. Consider experimental design; do not test all boards the same amount of time or even at the same temperatures!
PC board infant mortality ends after a month or three, so don't use the Weibull distribution to fit reliability after the end of infant mortality.
For more information about field reliability, what really happens, see my web site. Free samples.