Failure rate from degradation data

[vkp]

Hi,

I am trying to calculate failure rate (FITs) from degradation data. The data is in the form of degradation rates. Has anyone dealt with this problem? The key questions are:

1. How do you deal with the data if none of the units have failed?
2. How do you deal with units that are improving i.e. -ve degradation rates?
3. How do you deal with units that are essentially unchanged i.e. v.small degradation rates?
4. Can censoring be applied when you have zero fails or you are extrapolating time to fail?

Any input or references would be appreciated. Thanks.

[tarik]

Check out this reference: http://www.weibull.com/LifeDataWeb/degradation_analysis.htm
This type of analysis is implemented in Weibull++

[vkp]

Tarik,

Thanks for the link. It helps me with question 1. However, I am still not clear on how to deal with non-degrading units (zero degradation rates) or annealing units (-ve degradation rates).

Another question, is censoring applicable here?

[tarik]

2- How is the unit improving? Are you performing a growth program (if yes you can learn about the reliability growth analysis and RGA at http://www.weibull.com/relgrowthwebcontents.htm, http://rga.reliasoft.com/, http://www.reliasoft.com/RG/examples/index.htm)

3-You can use the same models referred to in the previous posts’ links, the difference is that you models will have parameters that describe a very slow degradation and failures that will happen way into the future.

4-When doing degradation analysis, you might have some units or even all of the units not failing (censored), so censoring applies in this case. You are measuring a quality characteristic (ex: ware, width, temperature, strength …) that describes the state of the product, based on the change of this quality characteristic you predict when it would reach a critical level that is considered a failure.

[vkp]

Tarik,

No, I am not doing a growth program.

Here's a brief description of the data. That might help clarify my question.

I have about 100 units that are tested periodically for a certain parameter (say, resistance). At time, t, their rate of resistance change is calculated and TTF for a certain allowable resistance is calculated. The data for the units is of 3 types:
1) 75% of the units show an increase i.e. positive TTF. They follow a lognormal distribution.
2) 10% show a decrease i.e. -ve TTF (These are the ones that probably should be censored but I am not sure what TTF to assign to them; if they are -ve TTF it gets ignored in failure rate calculations)
3) 15% show almost zero increase i.e. v.high TTF which is about 100 - 1000x of the units in group1. These artificially increase the sigma of the lognormal distribution and give high predicted failure rates. Should I be censoring these as well and if so, what TTF do I assign to them?

Thanks.

[tarik]

After using a degradation model and obtaining a list of TTF’s, I suggest you use the mixed Weibull analysis to fit a distribution, since you have different portions of your population behaving differently. As for the 10% that show a -ve TTF, those probably should be censored. In Weibull++, you can say that at the end of the test these units survived, or were censored, at the end of the test.