Operation Research Seminar : Nishant Mishra

Data-driven optimal policies for opportunistic mainteance of partially observable systems

Nishant Mishra, KULeuven

Abstract: With the onset of Industry 4.0, remotely monitoring a system's condition has become increasingly accessible. This allows for condition-based maintenance (CBM), whereby maintenance is performed based on the perceived condition of the system. Despite the value that can be derived from the implementation of CBM policies, the majority of OEMs still base their maintenance policies on two traditional maintenance approaches, namely corrective maintenance upon failure (CM) and preventive periodically scheduled maintenance (PM). In this paper, based on real data from an OEM, we develop optimal policies for systems that consist of components under CBM as well as under PM. Combining CBM and PM policies give rise to opportunistic maintenance (OM) where advantage may be taken of economic dependencies. We consider a setting that consists of a system with one continuously monitored component, maintained according to a CBM protocol, and a number of unmonitored components, maintained according to a PM protocol. The problem is modeled as a POMDP in two deterioration states, and we use a Bayesian approach to update the belief (the probability that the monitored component is in the warning state) at discrete decision epochs. We analytically characterize the structure of the optimal maintenance policy, and show that the optimal policy has three definite thresholds: one conditional on the time to the next PM action, and two on the belief state of the component. We also obtain extra thresholds, referred to as bundled thresholds, dependent on both the time and the belief state. Our research brings new insight to the bundling possibilities of maintenance actions and adds value for a decision maker regarding when Industry 4.0 based CBM may effectively, and cost efficiently, be applied in a dominant PM setting.

              Keywords: service operations, condition based maintenance, Bayesian updating, partial information, partially observable Markov decision process, preventive maintenance