Maintenance Optimisation for K-out-of-N: G System Using Deep Reinforcement Learning

In high-dimensional stochastic degradation environments, condition-based maintenance optimization for k-out-of-n systems is extremely difficult, making conventional dynamic programming techniques computationally unfeasible. This study uses a Gamma process degradation framework with scheduled inspections to formulate the maintenance problem as a Markov decision process. We propose a continuous-action Soft Actor-Critic (SAC) approach in which maintenance actions reflect the number of components that need to be replaced. We compare our SAC methodology with three discrete-action deep reinforcement learning methods: a Double Deep Q-Network (DDQN) for component quantity selection, a DDQN01 variation that provides binary maintenance decisions for individual components, and Proximal Policy Optimization (PPO). A threshold-based Monte Carlo method is also used for baseline comparison. When applied to a 3-out-of-5 feedwater pump configuration, the SAC technique outperforms threshold-based policies by about 3% on long-term average costs. Discrete-action methods produce undesirable intervention options, albeit having faster convergence rates. According to our sensitivity analysis, the benefits of deep reinforcement learning are greatest in situations with little redundancy or large-scale deployments, whereas traditional threshold rules are adequate for smaller, redundancy-rich systems. These results support the effectiveness and scalability of both continuous-action SAC and discrete-action deep reinforcement learning variants for optimizing multi-component maintenance, and they point to potential future uses in multi-objective cost structures and dependent failure scenarios.