| Abstract [eng] |
Age-dependent highly reliable systems provide small amount of statistical information and for that reason classical frequentist methods cannot be applied due to their asymptotical assumptions. However, Bayesian methods, due to their ability to naturally couple all sources of information (including expert subjective opinions) and not rely on asymptotic assumptions, are attractive approach to solve small sample problems in age-dependent reliability modelling. In this thesis Bayesian paradigm and its applicability were presented and general methodology to analyse previously mentioned problem was obtained. Methodology successfully was applied for two real data samples: failures in European natural gas grid and electrical Instrumentation and Control components. It was concluded that presented approach is able to easily investigate small samples in nonlinear age-dependent models. Also, analysis showed that different model goodness-of-fit approaches can provide different inferences and that sometimes it can fail due to nonlinearities and heteroscedasticity present in data. For that reason Bayesian posterior model averaging procedure were applied and concluded that it gives more reliable and better calibrated results than would be in one model case. Also adaptive Metropolis superiority over classical Metropolis – Hastings algorithm for highly correlated parameters and nonlinearities in model was validated. |
