Prediction Methods

Many people have heard of, or are familiar with various reliability prediction methods like MIL-HDBK-217, Telcordia SR332, etc.  These standardized handbook methods have widespread use in industry.  They are primarily applicable when making the assumption that the component failure rate is constant (at the bottom portion of the bathtub curve) and are thus generally applicable to most electronic components.  However, caution should be taken when using these prediction methods because there may be components for which this assumption is not correct including some electronic parts like electrolytic capacitors.  There are some handbooks that deal with mechanical parts but they also generally view the failure rates as constant for the time period of interest.

In working with a client recently, they had a reliability goal that they wanted to achieve and desired a reliability prediction to verify that the goal was achievable.  As their component parts list was reviewed, it became obvious that they had numerous parts that were subject to mechanical wear like an LCD touch screen, cable connectors, etc.  For the electronic parts, the goal could be achieved but the components subject to wear had to also be evaluated and integrated into the analysis.

It then becomes necessary to deal with such components that will experience wear individually and determine whether or not they are apt to wear out within the reliability goal period of interest (or product lifetime).  If it can be shown that the wear out occurs beyond the expected life of the product, then there is no problem.  This determination can be done through testing or other analysis methods.  If the component is likely to wear out within the expected product life, then decisions must be made regarding a maintenance strategy and the potential impact to warranty.

What has been your experience in performing predictions when you have components that can wear out?

An adaption of the Functional Safety standards IEC 61508 and IEC 26262 by the European Union brought a new life into slowly fading activity of reliability prediction. Both reliability prediction and reliability demonstration are now key parts of many product development programs, however despite phonetic similarity those two have little in common as well as the result they generate. 

While reliability prediction is an analytical activity often based on mathematical combination of reliabilities of parts or components comprising the system; reliability demonstration is based on product testing and is statistically driven by the test sample size.  Therefore the obtained results could drastically differ.  For example, a predicted system failure rate of 30 FIT (30 failures per billion (109) hours) would corresponds to a 10 year reliability of 99.87% (assuming 12 hours per day operation).  In order to demonstrate this kind of reliability with 50% confidence (50% confidence is considered low in most industries) one would need to successfully test 533 parts (based on binomial distribution) to the equivalent of 10 year field life.  Needless to say that this kind of test sample is prohibitive in most industries.  For example in the automotive electronics the test sample size of 23 is quite common, which roughly corresponds to 97% reliability with 50% confidence. 

The natural question is: how do you reconcile the numbers obtained from reliability prediction with the numbers you can support as part of reliability demonstration?

The answer is: I don’t believe that you can.

You can make an argument that reliability demonstration produces the lower estimate values.  Additionally the test is often addresses higher percentile severity users, thus the demonstrated reliability for the whole product population will likely be higher.  However, in most of the cases the gap will remain too wide to close.  This is something, which reliability engineers, design teams, and most importantly customers need to be aware of and be able to deal with as part of the product development reality.

What does the audience think? We’d love to hear your opinions on this.

Andre Kleyner

Below is a question from Paul Paroff asking:

“how to deal with these electromechanical components that do not have published MTBF values but do have rated life values in an MTBF prediction?”