SRC Forum - Message Replies
Forum: Reliability & Maintainability Questions and Answers
Topic: Reliability & Maintainability Questions and Answers
Topic Posted by: Reliability & Maintainability Forum
Organization: System Reliability Center
Date Posted: Mon Aug 31 12:47:36 US/Eastern 1998
Posted by: Bob Valerius
Organization:Northrop-Grumman Oceanic & Naval Systems
Date posted: Thu May 31 12:58:25 US/Eastern 2001
Subject: Modeling failure data
I have been asked to look at different methods for modeling failure data. On one of our programs, we model the reliability by using predictions from various sources. This program has been in test (sample size = 1) since 9/99. The approach we are taking is that when a component or subassembly fails, we replace its predicted failure rate with the actual. This of course almost always results in a lower reliability estimation. When the FRACAS procedure results in a corrective action being implemented, we then replace the actual failure rate with the predicted failure rate. This is how we report monthly on our reliability metric.
Does know of a reputable alternative approach? It has been suggested that since predicted failure rates are an average, to replace the predicted one with the actual, which is only 1 data point, is way too conservative. Any ideas or suggestions would be appreciated. Thank you very much.
Subject: Combining failure data
Reply Posted by: Bruce Dudley
Organization: Reliability analysis Center
Date Posted: Thu May 31 15:02:51 US/Eastern 2001
Combining Predicted Failure Rate With Empirical Data
Predicted failure rates for either parts or items can be combined with test or operating data to assess the combined effect. This is accomplished by mathematically combining the initial assessment made with reliability failure rate models based on past historical data with empirical data. This set combines the best pre-build failure rate estimate obtained from the initial assessment with the metrics obtained from the empirical data. Bayesian techniques are proposed for this purpose. This Bayesian technique accounts for the quantity of data by weighting large amounts of data more heavily than small amounts. The failure rate estimate obtained by prediction (PRISM or MIL-HDBK-217 or Belcore or other) forms the “prior” distribution, comprised of ao and bo.
If empirical data (i.e., test or field or operating) is available on the system or equipment or assembly or component, it can be combined with the best pre-build failure estimate using the following equation:
Lambda = (ao + a1 +----an) / (bo + b1 +----bn)
Lambda = best estimate of the predicted failure rate
Ao = equivalent number of failures of the prior distribution, (ao = 0.5)
Bo = equivalent number of hours associated with the reliability prediction, (bo=(ao/lambda p)
A1 through an = number of failures experience in the empirical data.
B1 through bn = equivalent number of cumulative operating hours in millions experienced in the empirical data