Integrated optimization of maintenance interventions and spare part selection for a partially observable multi-component system

Advanced technical systems are typically composed of multiple critical components whose failure cause a system failure. Often, it is not technically or economically possible to install sensors dedicated to each component, which means that the exact condition of each component cannot be monitored, but a system level failure or defect can be observed. The service provider then needs to implement a condition based maintenance policy that is based on partial information on the systems condition. Furthermore, when the service provider decides to service the system, (s)he also needs to decide which spare part(s) to bring along in order to avoid emergency shipments and part returns. We model this problem as an infinite horizon partially observable Markov decision process. In a set of numerical experiments, we first compare the optimal policy with preventive and corrective maintenance policies: The optimal policy leads on average to a 28% and 15% cost decrease, respectively. Second, we investigate the value of having full information, i.e., sensors dedicated to each component: This leads on average to a 13% cost decrease compared to the case with partial information. Interestingly, having full information is more valuable for cheaper, less reliable components than for more expensive, more reliable components.