Experience with Bayesian Reliability Measurement of Large Systems

Abstract

Experience with two methods of Bayesian reliability measurement is described. An aerospace subsystem was evaluated assuming continuous gamma-distributed component failure rates. Priors were developed by conventional reliability prediction methods based on handbook data. The ``strength'' of the prior was expressed in terms of variance about a predicted mean. Comparative evaluation was also made by a classical technique during a test program extending over 10 months. The Bayesian method was preferred though problems inherent in the method were apparent. More recently, a complex marine system was evaluated over a one-year period using a Bayesian formulation in which the failure rate is described by a discrete probability distribution with nonuniform cell widths. This technique avoids some of the operational problems of continuous formulations. Experience with Bayesian methods leaves little doubt of their utility as evaluation tools. The philosophical problems, however, remain as intransigent as ever.