Hi,
a customer has the following scenario:
They get from their customers motor testing data from a test rig (not field data). There they test under various conditions and always the full test time (which can vary between the tests, too). With this kind of data their customer expects them to estimate the failure probability in the field. So far so wierd 
|
test no.
|
stage
|
valve present
|
hours
|
test type
|
engine
|
|
5278
|
1
|
yes
|
465
|
Nardo
|
M9
|
|
5278
|
2
|
yes
|
465
|
Nardo
|
M9
|
|
6088
|
1
|
no
|
300
|
SHES+NC
|
M9
|
|
6088
|
2
|
yes
|
300
|
SHES+NC
|
M9
|
|
6117
|
1
|
yes
|
187
|
endurance
|
M9
|
|
6117
|
2
|
no
|
187
|
endurance
|
M9
|
|
6126
|
1
|
no
|
214
|
several
|
M9
|
|
6126
|
2
|
yes
|
214
|
several
|
M9
|
|
7291
|
1
|
yes
|
300
|
SHES+NC
|
M9
|
|
7291
|
2
|
yes
|
300
|
SHES+NC
|
M9
|
|
6107
|
1
|
no
|
392
|
Hot/ Cold
|
M9
|
|
6107
|
2
|
yes
|
392
|
Hot/ Cold
|
M9
|
|
7324
|
1
|
no
|
417
|
Hot/ Cold
|
M9
|
|
7324
|
2
|
yes
|
417
|
Hot/ Cold
|
M9
|
|
6917
|
1
|
no
|
416
|
SHES+NC
|
M9
|
|
6917
|
2
|
yes
|
416
|
SHES+NC
|
M9
|
|
6108
|
1
|
yes
|
562
|
Cmax/ Pmax
|
M9
|
|
6108
|
2
|
yes
|
562
|
Cmax/ Pmax
|
M9
|
|
1801
|
2
|
yes
|
314
|
Cmax/ Pmax
|
R9
|
|
1839
|
2
|
no
|
229
|
Hot/ Cold
|
R9
|
|
1839
|
2
|
no
|
400
|
Hot/ Cold
|
R9
|
|
1020
|
2
|
yes
|
435
|
SHES+NC
|
R9
|
|
833
|
2
|
no
|
388
|
SHES+NC
|
R9
|
|
1903
|
2
|
yes
|
357
|
spark end.
|
R9
|
|
1025
|
2
|
yes
|
279
|
Laquage
|
R9
|
Inn the data you see the test scenarious, two engines (M9 and R9). Their parts will be within the engines of the OEM who do the test with their engines (not in the field). Their parts include two rubber valves, listed as "stage". M9 has two, R9 only the second. A test run is a failure if at the end of the test at one valve is missing ("no" in valve present).
I've made some analysis (attached file), but I have some questions if this is at all an analysis one can trust.
1. As the test scenarios and the run times of the test are different does this impact the life distribution taking all observations into account? I guess yes.
2. Using test type as competing cause one could see that some test types are not included as there is just one or none failure in the data. This of course results in a huge confidence interval. If one would like at least to know when the a failure is to be expected they would have more information as of today. I guess this can be done already with that analysis, do I miss anything?
3. Is it a problem if they just detect a failure when the run is over? I compare it with repairable systems, and maintenance intervals, where you just detect there was a failure before that date, but not exactly to what time before that maintenance date. I'd expect that the uncertainty increases but probably not that much that the results cannot be used. Is that a too optimistic view?
How would you treat this problem if you cannot change the way the tests are run by the OEMs?
Thanks a lot,
Martin
