Sequential Failure Modes

[A Miner]

What is the proper method of analyzing sequential field failure modes with the following behavior? Failure mode A fails in an infant mortality mode, but does not disable or impair the system. Failure mode B CANNOT start until failure mode A occurs. Failure mode B then progresses as a low-cycle fatigue failure resulting in failure of the system. A Weibull analysis (including censored data) predicts 80% failures in a year, but actual experience after one year is 5%. Failure mode A is a limiting factor, since B cannot occur without it, yet failure mode B drives the Weibull results. A non-parametric analysis gives closer results, but cannot predict future failures.

[Pantelis]

What data do you have?

Scenario 1: You know when A occurred and when B occurred, and B can never happen until A occurs. If the modes are just sequential (mode B starts up only when A has occurred) but otherwise independent (i.e. it is independent of how long it took for A to happen) then one could fit a model (i.e. a Weibull) to time-to-event A and then another one for the time-to-event B starting from the time that A occurred. The system model would then be the probability that A occurs and then B occurs (i.e. similar to series configuration).

Scenario 2: All you know is the time to failure which is due to the B mode, and which presupposes that A has occurred. Your random variable is the time to event B assuming A has occurred. If you have sufficient failure data that spans the lifetime of the product fitting a Weibull (or even better a mixed Weibull) to it should give you a decent estimate. I am surprised to hear about your huge inconsistency. Can you share the data (or an example thereof) along with your assumptions and analysis?

[A Miner]

Scenario B is correct. All data is field data. We have no information at all on when failure mode A occurs since it does not impair system function. Failure mode A allows cyclical loading/unloading of a component that results in fatigue failure mode B. Prior to failure mode A these cyclic stresses are isolated from this component.

The basic problem with the Weibull analysis is that Failure mode B dominates the analysis in the 0 - 5% cumulative frequency range. Because Failure mode A is an infant mortality type failure, no failures exist beyond this, only suspensions. If you plot the cumulative failure rate against time, you see a curve that increases up to the 5 months service point then begins to decline. We have field data for unit with 2 to 13 months service and believe that failure mode A occurs within the first 2 months of service or not at all.

Regarding sharing of data, I will have to obtain permission to share.

[Pantelis]

I assume you meant to say Failure mode A dominates the analysis….

Then, you seem to have data and are able to model mode A. What you do not have (I am assuming) is an adequate model for mode B, since no failures have been observed for this (is this correct?).

To get an adequate model for mode B you may need to get additional data. In the absence of failure data an exponential model (constant failure mode) may be assumed for the mode or a one parameter Weibull. Both can be estimated from the suspensions.

The models for A and B may be then combined in a mixed Weibull model.

[Pantelis]

Typo Correction - '(constant failure mode)' should read '(constant failure rate)'

[A Miner]

No, Failure Mode B dominates. Since failure mode A does not impair the system other than allowing the cyclic stresses that result in failure mode B. We only have the combined failure times for failure mode A + B. We know that failure mode A precedes mode B based on failure analysis of returned units, and on engineering analysis of the design.

Failure mode B is a low-cycle fatigue type failure and Weibull plots of the data show a slope of ~1.8, which is characteristic of fatigue failures.

Failure mode A can occur almost immediately as evidenced by A+B failure times within 1 month of operation. Failure mode A appears to have a sharply decreasing failure rate typical of infant mortality. This restricts the population that will result in failure mode B. In fact, those that do not fail within 1-3 months appear to be "Perpetual Survivors".

If we were to see additional failures further out, I would expect to see the Weibull start with a steep slope then go to a very shallow slope less than 1. However, with no additional failures beyond a set age, I cannot prove this other than point to the massive number of units that have not failed.