A multi-state Markov model to separate the latent deterioration process from the maintenance effect
share
Maintenance optimization for naval ship equipment is essential in terms of national defense. However, the pure deterioration effect should be inferred in advance to compare the different maintenance policies. In other words, an exact inference of the effect of the deterioration and maintenance processes, which are mixed in an observed data, annual failure counts for example, is needed based on which policy selection could be performed. This paper proposes a new framework to retrieve the deterioration effect by separating it from the current maintenance effect based on the multi-state Markov model, using the annual engine failure count of 99 ships. The framework consists of multiple steps including imputation, transition matrix design, optimization, and validation. Hierarchical Gaussian process model is used for the imputation and the three-state Markov model is constructed to estimate the parameters regarding the deterioration and maintenance effect. Bayesian HMM model using a categorical distribution is applied to estimate the natural (deterioration) and artificial (maintenance) effect separately. Experiments showed the robustness of the estimated model parameters as well as the accurate recovery of the observed data under multiple settings, thereby confirming the credibility of our model. The results of this study could be employed to make a reliable system as well as to reduce a maintenance cost.
READ FULL TEXT