Fiabilité en fatigue des aubes des turbines hydrauliques dans un contexte imprécis

Abstract

The reliability of hydroelectric turbines is a complex function which depends mainly on the mechanical properties of the material and loading stress. These properties are affected by aging and operating conditions; therefore their original values considered during the hydraulic turbines design cannot be used along the useful life of the equipment. Hence, the need to the use of expert opinions to update these proprieties. In such cases, experts may rely on probabilistic theories or imprecise probabilities to formulate their opinions and update these properties. In paper # 1, we analyze how these theories affect the reliability calculation based on the FORM (first order reliability method) approach and having the Kitagawa-Takahashi diagram as a limit state. In this contribution we proposed an approach to extend the reliability calculation on variables expressed according to the imprecise probability theories. Also for the studied model, we highlighted that the variables expressed according to bounded distributions, reduce the model accuracy. To avoid this limitation, an approach that imitates unbounded distributions and respecting the physical behavior of the required variables has been suggested. The paper concludes that the modeling theories used to formulate expert opinions are equivalent and the priority should be given to opinions based on unbounded distributions. In order to formulate their opinions, the experts generally follow some elicitation techniques, appropriate to the studied subject. Examples of elicitation techniques proposed in the literature often control expert opinions by guiding them towards a consensus or to a specific choice. Otherwise, in the absence of these frameworks, experts formulate their opinions according to their own knowledge and according to their understanding of the subject, which may lead in some cases to disjoint or totally different opinions. This situation will be more complicated if experts have to predict data without any reference values. In paper # 3, we proposed and compared some elicitation techniques and obtained results showed that for domains where the required variable has a rich history, elicitation techniques with supports will be recommended in order to limit the variation between expert opinions. For multi-variable systems, we can have several experts available for the elicitation of each system input. In this situation, we wonder what the best way to combine these data is: before their propagation in the system model; or combining them after the propagation of each opinion separately in the system model. In paper # 2, we explored some parameters that can affect the difference between these two aggregation modes. In this sense we proposed the divergence metric

Date	6 Aug 2017
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Antoine Tahan (Supervisor) & Martin Gagnon (Co-supervisor)