GMW17626

Description / Abstract: GMW17626, 1st Edition, June 2016 - Best Practices for Weibull Analysis of Time-to-Failure Data

Purpose. This procedure provides guidance to GM and Suppliers on best practices for Weibull analysis of product life (time-to-failure) data. In this procedure, Weibull analysis refers to calculating the parameters of a Weibull distribution used to model time-to-failure data, and from that, to determine the reliability of the product.

Whenever test data is fit to a Weibull (or any other) distribution, that distribution is only an estimate of the distribution of the underlying population of parts from which the test parts were taken. The best practices given here are intended to improve the chance of getting a Reliability number that is as close as possible to the population’s true Reliability.

There are many different “settings” that can be used when doing Weibull analyses. Unfortunately, each decision typically leads to a different result. GM has preferred or required choices for some settings, and no official position (at this time) for others. The preferred/required settings are based on theory, simulations and other studies made by experts in the field, and observations from past analyses and studies within GM.

Applicability. The best practices given in this procedure are applicable when the following situations are both true:

• 
  
  The data being analyzed is observed times-to-failure for a product on a durability test, where the failures are due to a “wear out” type of failure mechanism (where the “damage” to the product builds up over time, eventually leading to a failure). The data may also include:

a. The amount of testing performed on products that have not yet failed – these are referred to as right censored or suspended items; and

b. A range of times within which a part failed when the exact failure time is not known – this is called interval censoring, or left censoring if the left-most point is at the start of test.

• 
  
  The portion of the Weibull curve of greatest interest is at the low end of the distribution, where there is a low percentage of parts that have failed, which means high part reliability

The practices given here may not be the best choices under other situations, such as analyzing customer usage data (where the upper end of the distribution is usually of greater interest), or strength data such as material yield strength (which usually has much higher Weibull slopes, or which may be better fit using a different distribution). Therefore, the practices given here should not be used blindly for other situations.