Reliability Growth methodology (Duane, CrowAMSAA) - Test Planning

[doops]

Hello,

I am trying to implement Reliability Growth methodology (Duane, CrowAMSAA) in my company to help with field test sizing and test time.

I would like to know any other companies that use this methodology for test planning. Identifying companies and industries which use reliability growth models for test planning would help me sell this idea.

Thanks.

 

[Sturmkind]

Hi, doops!

All automotive and aerospace companies use this methodology or a variation thoughout the product development and product realization cycle. Specifically, automotive safety restraints (seat belts, air bags) have a target of 5 '9' reliability meaning 0.99999 reliability @ either 0.90 or 0.95 confidence. depending upon whether it is Design Reliability (generally .99999 @ 0.90 confidence with zero failures) or Process Reliability (generally .99999 @ .95 confidence with zero failures).

The initial DFMEAs & PFMEA's start with higher Risk Priority Numbers (RPN) since some performance test results are considered unknown or unproven so the detection and occurrence values will be higher, regardless of severity. Initial testing of small quantities without failure may allow the detection and/or occurrence values to begin to be reduced. A single failure should result in a design change of either product, process, or both. At program Milestone validation points, test quantities are increased as Start Of Regular Production approaches and successful testing is subtracted from the production life goal. After all, with vehicle life at ~20 years demonstrating .99999 @ .90C 230,259 OK tests. This quantity is feasible for something like number of landings on an aircraft carrier, but not so good for the carrier itself.....

Typical companies that use this are all OEMs, Autoliv, Key Safety Systems, Takata, TRW, and so forth.

 

[doops]

Hello Sturmkind!

Thank you so much for replying!

Just out of curiosity, would you please elaborate a little bit more on the reliability growth testing? Is the growth plan through several individual test phases, and they use reliability growth models to tie all the tests together?

For example, after test phase I, the article has 100 cycle MTBF. Fix failures. Phase II, initial MTBF is 150 cycles. Test. After Phase II, MTBF is 200 cycles. Fix, and continue until meeting reliability targets?

Is the reason for using reliability growth, because one single zero failure test would yield outrageous sample sizes and test times?

Using binomial equation, I calculate needing 23,026 samples with 10 lives to get R 99.999% & C 90% with zero failures.

Thanks again!

 

[Sturmkind]

There needs to be a growth plan associated with each response variable expectation.

Using a seat-belt system as an example there will be several key responses:
A) Ability to cycle 'x' times full extraction / retraction without exceeding 'y' Nm force in 'z' seconds.
B) Lock-up speed of the system in 'x' milliseconds upon impact at 'y' Nm force.

Those would be performance related.

C) Design to restrain a 95% adult with minimum neck/torso Nm force upon impact of 'y' Nm.
D) Lock-up speed of the system in 'x' milliseconds upon impact at 'y' Nm force with full system retention integrity. Meaning, upon impact the mooring bolts don't tear out of the floor, the seat-belt works as intended with the air bags, the seat integrity is maintained, the occupant is restrained without undue injury, and so forth.

Cyclic testing of A & B can start and continue as C & B are validated. C & B will have much smaller test sizes and a test plan with a flatter angle and larger confidence bands than A & B because the tests are destructive, and very expensive.

Does that help?

 

[doops]

I think I understand. I appreciate you taking the time to answer my questions.

Is the Duane, CrowAMSAA or another growth model used in the validation planning and tracking of reliability?

Thanks again!

 

[Sturmkind]

Statistical Steven may disagree with me but they are variations on a theme. If the results are within 5% of another then they are considered equivalent. The log/log model should be fine.