Maintenance optimization of multi-unit balanced systems using deep reinforcement learning

Abstract

Balanced systems are extensively utilized in various engineering fields. However, the study of maintenance optimization for balanced systems under dependent competing risks is very limited, and the existing method for orthodox multi-unit systems cannot solve the maintenance problem of balanced systems. Thus, we present a novel maintenance optimization method for systems with multiple balanced units. The units in the system are assumed to experience multiple failure processes due to continuous degradation and external shock damage. The system failure occurs in the following situations: (1) the deterioration levels of units on the symmetric positions exceed critical thresholds. (2) the deterioration level of any unit reaches the failure threshold. (3) the shock magnitude is greater than the hard failure threshold. By employing a customized deep reinforcement learning algorithm, we derive the optimal maintenance policy under a fixed planning horizon. Notably, our approach differs from previous research on maintenance optimization for balanced systems, the proposed method avoids the discretizing of the deterioration state. This significantly enhances computational efficiency, particularly when applied to multi-unit balanced systems. Numerous numerical studies illustrate the effectiveness of our proposed maintenance optimization method for multi-unit balanced systems.