An integrated approach to a condition based maintenance policy and applications

Unexpected system failures pose a significant problem in human safety and health care applications, service and manufacturing sectors, national infrastructure (nuclear power plants and civil structures), and national security (military operations). The main challenges associated with unexpected failures are related to characterizing the failure uncertainty and the stochastic nature of the degradation processes. An accurate failure time prediction and a reliability assessment are necessary if the appropriate maintenance resources (personnel, tools, spare parts, etc.) are to be assembled. For this reason the thesis presents a mathematical framework for integrating degradation-based sensor data streams with high-level logistical decision models. To achieve this goal, a software has been realized in order to simulate a discontinuous operational scenario (such as aircraft operations) in which two different maintenance policies were applied, a scheduled and a condition-based one. The former refers to a typical maintenance policy, in which no prognostic data are available, so that maintenance is scheduled basing only on prior knowledge of components’ failure behavior. The latter approach, instead, implements the information given by prognostics in order to fully exploit the component’s residual useful life and reduce the lead time to deliver spare parts. The last change is achieved through a revision and a modification of the entire supply chain model in a Just-In-Time-like perspective: thanks to a more precise knowledge of the time to failure, spare parts can be stored in depots so to be in the maintenance zone just before they are needed. Thus, it is possible to move these parts to higher level depots, where hold stocking costs are typically lower. As for prognostics, it has been made possible through the realisation of a RUL estimation algorithm. It is to say that many techniques have been found in literature, but none of them faced the prognostic problem with the aim of finding a closed form for RUL estimation The most promising predictive algorithm, among those developed before this work, turned out to be a Bayesian estimator based on the degradation pattern of the monitored component, under the likely assumption of exponential shape of such pattern. This algorithm has been the starting point for the one developed in this work. Leveraging on Bayesian probability theory, the up-to-date RUL probability density function of the component is evaluated at each time step, starting from the prior knowledge of the component’s residual life, a stochastic parameter that is evaluated from experimental tests always done before commissioning. The ________________________________________________________________ information about RUL prediction was then used to define the optimal moment at which scheduling and performing maintenance. These values were found through an objective function optimization that took into account the main drivers associated to condition-based maintenance decision making process. Furthermore the opportunity to introduce CBM (condition based maintenance) concepts based on prognostic into a cracked railway axle management is investigated. The performances of two different prognostic algorithm are assessed on the basis of their RUL (remaining useful life) predictions accuracy. The CBM approach is compared to the classical preventive maintenance approach to railway axle maintenance management based on expensive and regular NDT. The effect of monitoring frequency and the monitoring infrastructure size error is assessed as well.