A Bayesian approach is often applied when updating a deterioration model using observations from inspections, structural health monitoring, or condition monitoring. The observations are stochastic variables with probability distributions that depend on the damage size. Consecutive observations are usually assumed to be independent of each other, but this assumption does not always hold, especially when using online monitoring systems. Frequent updating using dependent measurements can lead to an over-optimistic assessment of the value of information when the measurements are incorrectly modeled as being independent. This article presents a Bayesian network modeling approach for the inclusion of temporal dependency between measurements through a dependency parameter and presents a generic monitoring model based on the exceedance of thresholds for a damage index. Additionally, the model is implemented in a computational framework for risk-based maintenance planning, developed for maintenance planning for wind turbines. The framework is applied for a numerical experiment, where the expected lifetime costs are found for strategies with monitoring with and without dependency between observations, and also for the case where dependency is present but is neglected when making decisions. The numerical experiment and associated parameter study show that neglecting dependency in the decision model when the observations are in fact dependent can lead to much higher costs than expected and to the selection of non-optimal strategies. Much lower costs (down to one quarter) can be obtained when the dependency is properly modeled. In the case of temporally dependent observations, an advanced decision model using a Bayesian network as a simple digital twin is needed to make monitoring feasible compared to only using inspections.
Proposal for a Time-Dependent Dynamic Identification Algorithm for Structural Health Monitoring
