A Simulation-based Bayesian Approach for Parameter Estimation with Model Misspecification Analysis using ALT Data
Résumé
Accelerated life testing (ALT) is a standard approach to gather information on the failure times of highly reliable devices. In ALT, devices are exposed to higher levels of stresses (e.g. higher temperature, voltage, pressure) to produce failures more quickly and, hence, reduce the cost and length of tests. The collected data obtained at these levels are then analysed and extrapolated to obtain the reliability function at use stress level. However, before extrapolation, one needs to select an appropriate distribution that better reflect the variability of the times to failure and an acceleration law. In practice, often several statistical distributions could be used either from statistical or physical considerations. Simpler models are always preferred as they offer many advantages to practitioners. Hence, the primary aim of this work is to investigate the effects of model misspecification specifically when the threeparameter Weibull distribution is incorrectly specified as one of the following non-nested models: the Birnbaum- Saunders and the lognormal commonly used when the failure is caused under cyclic loading (fatigue). To estimate the parameters of the proposed ALT models, an efficient variant of approximate Bayesian computation algorithm called ABC-NS is used.