Effect of suspensions on weibull parameters and life

[Suraj]

Hi ,

I am working on the analysis of life data for a component where in there are a large many suspensions. Could you please help me in understanding what the effect of suspensions on the weibull parameters.
I would like to understand the effect in the below 2 scenarios

case 1

eg we have a 100 sample size of which only 10 % have failes while suspensions have life less than the failure life for the 10 samples.
while generating the weibull parameters do I include these suspensions.

case 2

eg : I have samples which have suspensions having life more than the failure life of failed samples.

what is the best condition for including suspensions in the analysis? and what would be effect for both the above cases scenarios,
Would it be right not to include any suspensions atall to be on a conservative safe region.

Kindly guide

Regards

Suraj

[Pantelis]

Yes! You include them.

As an example lets say you want to find the average life span of humans based on a group of 100 that you are observing. If one died at 40 and one at 30 out of the 100 and the rest (98) are still alive – would you say that the average life span is 35 (based only on the observed deaths)?

On the question of would it be right not to include them and be conservative? Absolutely not…. You are not being conservative in your analysis – just wrong.

[Bende Sandor]

Hi,
I have similar but more extreme problem: what information can we gather from a sample where everybody is suspended?

[Pantelis]

Well… all you know is that by time X nothing failed. This in turn can be useful in determining an estimate of the reliability at time X. Now what I think you are asking is how one obtains a model (i.e. fit a distribution to the data) for interpolation/extrapolation purposes. Trying to fit a standard Weibull model is not possible because there are no failures in the data set, and thus one cannot fit a model that assumes a time varying failure rate to it. One option is to fit an exponential model to it – and assume a constant failure rate. This will provide results, however one should be careful in extrapolating past time X – especially with the constant failure rate assumption. A better option is to utilize a one parameter Weibull model. This would require that you assume a value for the beta parameter – from past designs, engineering knowledge, etc.

Hope this helps.