Dr. Ipek Dursun (Linkedin) Nominated Maintenance Thesis Award
Maintenance optimization and spare parts management in
data-integrated environments
Capital goods are crucial for the continuation of production and services. Many business operations are dependent on a functioning capital good. The efficiency of maintenance operations is important to keep these systems functioning. Maintenance service providers (SPs) are responsible to provide maintenance services and spare parts to their customers according to service level agreements. The first part of this dissertation (Chapters 2 and 3) is on maintenance optimization policies under the so-called time-to failure model uncertainty and the second part of this dissertation (Chapters 4 and 5) is on spare parts management by using advance demand information (ADI) from the point of view of an SP of capital goods.
In Chapters 2 and 3, we consider newly designed systems with a fixed lifespan. In all systems, the same critical component occurs that is subject to random failures. The component can be replaced preventively to avoid a costly failure. This component always comes from either a weak population or a strong population. This is known as population heterogeneity. The true population type is unknown to the decision-maker, but there is a belief with respect to the probability of having a weak population. This belief is updated with a Bayesian approach by using the data collected over the lifespan of the system.
In Chapter 2, we focus on a single system. We build a partially observable Markov decision process (POMDP) model to find the optimal replacement policy that minimizes the total cost over the lifespan of the system. It optimally balances the trade-off between exploration, i.e., the cost of learning the true population type (via deliberately delaying the preventive replacement time), and exploitation, i.e.,
the cost of maintenance activities. We generate insights on the optimal policy and its structure and we compare its performance with existing heuristic approaches from the literature, namely a myopic policy (i.e., a heuristic that does not consider exploration) and a threshold heuristic (i.e., a heuristic that considers exploration). In the numerical analysis, we observe that the true population type is learnt much faster under high exploration. Additionally, the myopic policy and the threshold policy are up to 23.6% and 5.8% costlier than the optimal policy, respectively.
In Chapter 3, we consider multiple identical systems. We investigate the effect of socalled data pooling by updating the belief with the data collected from all systems. We build a discrete-time POMDP model to find the optimal replacement policy which minimizes the expected total cost throughout the lifespan. We compare the cost per system under the optimal policy with the cost per system under two benchmark heuristics that follow the single-system optimal policy with and without data pooling, respectively. We investigate the effect of joint optimization and data pooling, and the number of systems on the cost per system. In our numerical experiments, we show that the cost reduction relative to the worst benchmark heuristic (i.e., the benchmark heuristic without data pooling) can be up to 5.6% for two systems, and this increases up to 14% for 20 systems.
In Chapter 4, we study multiple technical systems that are supported by a local stock point. We consider a single critical component that occurs in each system and is subject to random failure. After a failure, a replacement takes place. Signals for possible failures are generated by predictive models, which constitute ADI for the spare parts inventory. However, signals might be imperfect. We assume a periodic-review replenishment policy. We formulate a Markov decision process model to find the optimal inventory replenishment policy that minimizes the longrun average cost per period. We investigate the effect of precision (i.e., the fraction of signals that are true positive), sensitivity (i.e., the fraction of failures for which a signal is generated), and the demand lead time (i.e., the time between the signals
and failures) on the optimal costs. We show that a significant cost reduction can be obtained for a moderate value of precision. For the sensitivity and demand lead time, you always need high values in order to get a significant cost reduction.
In Chapter 5, we consider an SP that organizes maintenance visits and spare part shipments to a customer when a failure code is reported. This code constitutes a form of ADI. We formulate a mixed integer linear programming model to find the optimal set of spare parts that will be sent to a customer site to resolve the failure. The optimal set minimizes the total costs consisting of shipment costs, costs for the return of parts that are not needed, and costs for a required second visit if the set does not contain all parts to solve the failure. We derive analytical results for the structure of the optimal policy. We compare the policy generated by our model to existing benchmark policies. In an extensive numerical study, we observe that one specific benchmark policy cannot find the optimal policy for all problem instances. The best benchmark policy is on average 12.2% costlier than the optimal policy.