Meta‐RL Based Adaptive Selective Maintenance Scheduling for Multi‐State Systems in Dynamic Environments

Selective maintenance, as a prevalent maintenance policy, has been widely applied in various engineering and military scenarios. Most existing studies assume a stationary environment, under which degradation dynamics remain fixed and fully known. In practice, however, the operating environment evolves dynamically over missions, altering component degradation intensities in ways that cannot be directly observed at the decision epoch. This paper formulates selective maintenance under dynamic environments as a partially observable Markov decision process (POMDP), where the environmental effect is modeled as a latent context evolving over missions. A novel meta‐reinforcement learning (meta‐RL) approach is proposed to acquire adaptive decision‐making capability from multiple representative fixed environments, enabling the inference of optimal selective maintenance policies in dynamic settings. Two cases validate the method: a two‐component system demonstrates rapid recovery after latent context switches with near‐optimal decision quality, and a 14‐component coal transportation system confirms long‐term adaptation to seasonal variations. Results show that the framework enables effective and rapid adaptation for selective maintenance under dynamic environments.