Um modelo identificável para dados de degradação quando unidades operam em condições distintas

Abstract

Information on the failure times of highly reliability products, that are unlikely to fail traditional life tests or accelerated life tests, can be obtained from degradation experiments. Ferreira et al. (2012) proposed a linear degradation model in which the degradation rate depends on categorical covariates. In this article, theoretical aspects of this model are studied and it is shown that such a model is unidentifiable. The effect of the model non-identifiability in the posterior sampling via Markov chain Monte Carlo methods is discussed. Modifications in such a model structure are proposed, obtaining an identifiable model. It was implemented under the Bayesian approach. The results obtained in the simulated data studies performed showed that the proposed model produces reasonable estimates for time-related quantities until failure.