Partially observable Markov decision process-based optimal maintenance planning with time-dependent observations

Abstract

The growing technological capability for real-time condition monitoring (CM) of industrial equipment has spurred significant interest in methods for optimal maintenance planning using CM signals. Existing approaches for maintenance policy development consider degradation to be either fully or partially observable. For the more general case of partial observability, it is usually assumed that the relationship between the underlying degradation process and the observed condition is time-invariant. In this paper, we address this major shortcoming by modeling observed CM signals through an underlying failure process wherein the linkage is time-dependent piecewise linear with jumps, and then utilizing a Partially Observed Markov Decision Process (POMDP) to determine the optimal maintenance strategy. We investigate the structure of the policy and show that, under certain conditions, a control-limit policy exists, i.e., a belief threshold exists beyond which the optimal action is to preventively maintain the unit. Finally, we present a case study based on battery resistance data and demonstrate that our modeling procedure offers a maintenance policy that is superior to those from other competing models.