Statistical Analysis on Limited Fatigue Data

[bcats4life]

Hello,

I am looking for some statistical methods that may be applicable to a study that I am performing. I am currently performing some vibratory high cycle fatigue (HCF) testing on some parts in a 1F mode. I need to demonstrate that the parts have a certain amount of HCF strength, for example 10 ksi. So far I have tested 9 parts and all parts have demonstrated to “runout or pass” at 15 ksi or above. The issue that I have is that I have been unable to break the parts in the mode of interest, unfortunately changing the mode is not an option. If I were able to break the parts, that would tell me what the actual HCF strength is. Instead, once I get to levels between 20 ksi – 35 ksi, varies for each part, the part starts coupling and I can no longer run in the mode of interest.

Enough background, here is my statistical question. The two options that I think I have with the current data is to perform some type statistical analysis that analyzes the go no-go data (I have 9 parts that pass and 0 parts that do not, what does this tell me statistically) or I could do some type of analysis that uses the known minimum HCF levels that I have determined for each part (for example, I know that 4 of the parts that have a strength of at least 15 ksi, 3 parts have a strength of at least 20 ksi, 1 part has a strength of at least 25 ksi, and one part has a strength of at least 35 ksi). If anyone has any ideas on some type of statistical analysis that I can do to demonstrate that I have plenty of HCF margin (compared to my 10 ksi requirement) which would give me confidence when these parts start in production that I will always meet my 10 ksi limit, please feel free to share. If there are other types of options that may be applicable for my study, please also share those ideas.

Thank you.

 

[PSE]

For getting some useful data out of smaller sample sizes (I am no expert on this). There is a lot there and I think you could use a plot to determine if a population % would actually fall into true design failure limits.

http://www.weibull.com/

I think off hand your use of the "coupling" strength is a good idea for obtaining variables data, to me always preferrable to attribute data. It looks like you have identified an (analytically acceptable) "failure" mode which does not involve component destruction.

Regards,

 

[bcats4life]

Is a Weibull analysis reasonable to use when I have step-stress data and failures? In other words, I have performed HCF testing on various components and found that the HCF strength reaches anywhere from 10 ksi to 30 ksi before failure occurs. Can I use a Weibull to accurately analyze this data or does the x-axis on a Weibull need to be time based? All Weibull examples that I have seen to this point have used some unit of time on the x-axis.

Thanks

 

[IRstuff]

Don't see why not. Weibull's seminal paper doesn't really discuss in terms of time:

http://www.barringer1.com/aug01prb.htm

Lots of stuff on the Barringer website:

http://www.barringer1.com/wa.htm

http://www.barringer1.com/Papers.htm

TTFN

FAQ731-376