Failure of emergency diesel generators: a population analysis using empirical Bayes methods

This paper presents analysis of failure data from a large population of emergency diesel generators (EDGs) at nuclear power plants. Empirical Bayes methods are used to estimate the population distribution and the individual failure probabilities from EDG's failure to start and load-run data over 4 years for 195 EDGs at 63 plant sites. Each diesel failure probability is treated as being a sample value from an underlying population distribution. The statistical characteristics of the failure probability distributions (mean, variance, standard defiation and others) are estimated. The estimates of the failure probability distributions are useful for evaluation of EDG reliability and can be used as prior distributions and in uncertainty propagations in probabilistic risk assessments (PRAs).