System vs Component data

[Alypius]

The situation is that we wish to test a machine until mechanical failure, replace the component, continue operation until another failure occurs regardless if it is the same component that fails, replace that component, test until mechanical failure, etc...
By doing this we hope to plot a Weibull for the machine once enough failure data is obtained for each component of concern . I anticipate a bent Weibull, indicating multiple failure modes. The question is, can I assume that this system curve is equivalent to the "worst case" union of all the component Weibulls or would it be the product of the reliability of each contributing component? What is the relationship between the system reliability and what can be expected of the components. Obviously, the goal here is to determine replacement frequency of component parts in the shortest amount of time with the least test cost.

[RS Support]

I assume what you are analyzing is the times to failure of each component. In this case (and if you use competing failure modes analysis see http://www.weibull.com/LifeDataWeb/competing_failure_modes.htm ) you will get a curve as indicated, and your anticipation is correct. Also as mentioned, the system reliability will be the product of the reliabilities of the components (i.e., a series reliability system - and assuming component independence).

Now if you need replacement frequency for each component, doing the analysis individually for each mode will give you a Weibull for each component that you can utilize for further analyses (including ways to come up with the optimum replacement period for each single component -- see discussion at http://www.reliasoft.com/newsletter/2Q2000/preventive.htm --)

[Massimo_Barbi]

I partecipated at Weibull++ and Block Sim training course, but something of fundamental concept is not clear for me yet. My customer (Ford) wants that I use Weibull++ software; but my problem is: I have to analyze a machine with for example 100 different components, and the data derived from machine monitoring are like these. If, for example, the observation period is 1 year, I had, for example, 30 failures of 30 different components; 1 failure each component. So How can I use Weibull++ ??

All the examples, showed in the reliasoft training course, they concerned about a test where we had, for example, 10 identical components, and we analyzed the time failed data. But my situation is completely different: I have a machine and not a test in laboratory, I have a lot of components but none identical, I have only one failure for some components but only one for the same component.

Therefore I bought Weibull++ & BlockSim Software, but I can't use it till I will understand this concept.

Yours faithfully
Massimo Barbi

[RS Support]

Massimo,

There is a specific methodology for the FORD Machine Tool initiative you mention. Weibull++ MT and BlockSim MT are designed to address the methodology. A simple example can be found at http://www.reliasoft.com/Weibull/mt/mtexamples.htm.

When you mention training, which course are you referring to?

There are specific courses just for the FORD initiative that are offered by AIQ (http://www.aiqusa.com) in Detroit. AIQ is also the designated entity for help regarding this initiative. You can call them at 734-421-6413.

If you need further assistance please feel free to contact our help line at 520-886-0366, or email us at Support@ReliaSoft.com .

[Dr Chuck]

I looked at your example from Meeker and Escobar, and the Weibull fit to the Mode 1 data was NOT very good. Given the nature of the failure mechanism(voltage spike), I would have expected an Exponential distribution if the components were homogeneous and the voltage spike process was Ergodic. I think you should have considered a model which included a probability distribution over the components' spike susceptibility thresholds, and a "Frailty Removal" model for the "suspended" components.

Dr Chuck

[Pantelis]

Dr Chuck.

Let me begin by saying that the example in question is presented to illustrate a general methodology on analyzing competing failure mode data, and is not intended to categorize a particular failure mode, physical phenomenon, or circumstance.
Secondly and with regard to your comments, what is your assertion ‘on being a bad fit’ based on? An exponential fit is worse that the Weibull fit for this data set!
With regards to your last comment, one could have used a different approach, a more complex model, or additional information if such information was available and if more data was available.

Pantelis Vassiliou
ReliaSoft Corp.