Integrating probabilistic risk assessment with the partially observable Markov decision process to consider risk constraints in maintenance optimization

This paper is focused on the maintenance optimization of safety-critical systems that are under constraints on the system risk during their operation. The proposed method for addressing this problem integrates probabilistic risk assessment for risk modeling and updating within the partially observable Markov decision process (POMDP) to ensure that pertinent constraints on the system risk are properly considered in the developed maintenance policy. Although POMDP has been widely used in maintenance optimization, existing studies based on POMDP have not explicitly considered such risk constraints when solving the optimization problem. Compared with existing heuristic and risk-based methods for maintenance scheduling, the proposed method leverages POMDP to provide a condition-based and predictive maintenance policy. The new method is demonstrated and compared with two baseline methods for risk-based maintenance through a nuclear power plant system. The results show that the new method achieves more than 50% cost reduction compared with the baseline methods. • Maintenance optimization under system risk constraint. • Integration of PRA and POMDP. • Deep reinforcement learning for solving maintenance optimization. • Case study shows substantial cost reduction compared with baseline risk-based methods.