From The Editor

The Trees and the Forest

What are the burning issues in Reliability? In this issue's opinion piece by Kam Wong, you can see we're still arguing over the definition of random failure and the validity of the exponential distribution. We have also argued over the accuracy of MIL-HDBK-217, the relative merits of statistical vs. deterministic models, the validity of the "bathtub" curve and the reliability of plastic devices. We have in the past had arguments over the difference between hazard rate and failure rate, the use of Bayesian statistics, and the concept of reliability growth vs. doing it right the first time. In my opinion, we have spent a great deal of energy discussing details which would have been better spent considering the big picture. While concentrating on the trees, we may have lost sight of the forest.

I submit that most of these issues are arguments of viewpoint. Some examples:

A parts vendor needs deterministic models to design his parts, but a logistician needs statistical models to create his spare parts lists. These are never quite right, of course, but the alternative is to start with a random number. "In the land of the blind, the one-eyed man is king."

And good old MIL-HDBK-217 is another one-eyed man in the land of the blind. It provides a synthesized benchmark based on consolidated experience. If you have nothing and need a planning figure, it is there for your use. If you need better data, you'll probably have to run some tests. I feel that Charles Leonard had a valid complaint about putting million dollar cooling systems on Boeing airplanes based on a prediction he did not believe. But wouldn't it have been worthwhile to run a million dollars worth of flight tests with the cooling system turned off? It certainly would have decided the issue. Without test data, you can have any prediction you want by selecting among the failure rates in the GIDEP telephone books, or you can use MIL-HDBK-217 as a standardized reference based on a lot of data and models that someone is at least trying to keep up to date. You may not really need a prediction at all, but if you do, 217 may be the most sensible approach.

How about the "roller coaster" curve as a replacement for the bathtub? Suppose it is true that failure rates follow a roller coaster depending on the defects found in the product. An empirical roller coaster plot would be useful in determining what needs to be corrected. Removal of all defects would be shown when the ideal profile (guess what? the traditional bathtub curve) was achieved. For planning purposes, there is no way to predict the shape of a roller coaster curve, and it only shows a situation needing correction.

How about exponential vs. Weibull? If you have a mixed population of short lived parts which you replaced only at failure (e.g. light bulbs*) you will eventually have a constant failure rate** (exponential distribution of failures). If you believe in endless burn-in, there will come a point when the failure rate changes slowly enough to be considered constant. Automobiles and mechanical components often need Weibull analysis. For a great many things the exponential assumption either applies or is close enough to warrant using the simpler analysis methods it permits.

So use whatever arrow fits your bow. Meanwhile, instead of arguing about viewpoints, let's do some thinking about our role in life. When CAD data automatically feeds CAM machines, where does reliability get done? What does the reliability engineer do in the information age, and how does he do it? Are we wasting effort on bureaucracy (ISO 9000 compliance, etc.) which should be spent on technology? Have we got a handle on software? Is there anything significant we can do to improve human reliability? Are we part of the solution or part of the problem?

Don't be too quick to answer that last question.

* O.K., so light bulb failures follow a normal distribution. A Weibull with beta equal to 3.44 is close enough.

** Use hazard rate if it makes you feel better.

Anthony Coppola
Editor

Note: I am willing to bet Kam Wong will have some comments on this editorial; how about the rest of you? Readers are invited to reply by letter to the editor or by submitting an opinion article. Those inputs of general interest may be reprinted in future issues of the RAC Journal, unless the author requests otherwise. I expect to get some flak about the examples cited, but I do hope our readers will give some serious thought to the big picture issues and share the results.

Back to
Journal Index

User Guide

Products
and Services

Contacting
RAC

*Related
*WWW Sites

Home

From the Reliability Analysis Center Webserver. Comments or questions can be submitted using e-mail or telephone at (315) 337-9933. (Last update: December 1, 1993, jmc)