I’m an engineering student and currently working on a project dealing with the availability of a complex technical system. I have access to about 50 datasets for the components of the system. My problem is that the available data is not actual “failure” data but only “preventive replacement” data. I’m looking for a method to approximate component reliability with the available data. I’ve so far only found methods where there is at least one actual failure. I would greatly appreciate any feedback on this subject.
Thank you very much in advance.
Take a look at using Weibull analysis. The non-failed items are “suspensions”, which Weibull analysis handles. Ideally you should have some failures in the data set to use the Weibull approach. However, if there are very few, there is an approach called WeiBayes that assumes a Weibull slope (beta), thereby allowing a Weibull analysis to be applied (see http://www.bobabernethy.com/case_study_Wang_Weibayes.htm).
Thank you very much for your help,
I now have Dr. Abernethy’s “New Weibull Handbook”. I understand I have to use the one parameter Weibull distribution (WeiBayes) when my data consists of supsensions only. I can estimate a shape parameter. I would just like to compare it to some generic values. The only values I could find were posted by Barringer and Associates, Inc. Do you know any other “public” sources?
Concerning the estimation of eta, I have to partially differentiate the likelihood function. The resulting function is described in Dr. Abernethy’s book. Do you know a source where the partial differentiation is described step by step?
I would recommend that you give Paul Barranger a call/email and request permission to use. I know of no other “publis” sources.
Regarding “partial differentiation”, I do not have a specific reference that I can recommend (that provides “step-by-step” instruction); however, there are many good math/calculus books that describe this. I would go to a university library and browse some to see what fits your needs.
Yes, this is correct. Otherwise, you have to assume a failure when none really exists in order to prevent a divide by zero error.