Thesis for the Degree of Licentiate of Engineering

Maintenance is a source of large costs; in the EU the maintenance costs amount to between 4% and 8% of the total sales turnover. Opportunistic maintenance is an attempt to lower the maintenance cost by considering the failure of one component as an opportunity to replace yet non-failed components in order to prevent future failures. At the time of failure of one component, a decision is to be made on which additional components to replace in order to minimize the expected maintenance cost over a planning period. This thesis continues the work of Dickman et. al. [9] and Andréasson [1] on the opportunistic replacement problem. In Paper I, we show that the problemwith timedependent costs is NP-hard and present a mixed integer linear programmingmodel for the problem. We apply the model to problems with deterministic and stochastic component lives with data originating from the aviation and wind power industry. The model is applied in a stochastic setting by employing the expected values of component lives. In Paper II, a first step towards a stochastic programming model that considers components with uncertain lives is taken by extending the problem to allow non-identical lives for component individuals. This problem is shown to be NP-hard even with time-independent costs. We present a mixed integer linear programming model of the problem. The solution time of the model is substantially reduced compared to the model presented in [1]. In Paper III, we then study the opportunistic replacement problem with uncertain component lives and present a two-stage stochastic programming approach.We present a deterministic equivalent model and develop a decomposition method. Numerical studies on the same data as in Paper I from the aviation and wind power industry show that the stochastic programming approach produces maintenance decisions that are on average less costly than decisions obtained from simple maintenance policies and the approach used in Paper I. The decomposition method requires less CPU-time than solving the deterministic equivalent on three out of four problems.