Ontology-driven cloud dynamic Bayesian network model integrated with deep reinforcement learning for dynamic maintenance decision-making

The efficacy of maintenance decision-making for engineered systems is influenced by the actual state of the equipment and the corresponding degradation mechanisms and failure modes. The unpredictable evolution of time-varying degradation states, combined with the uncertain failure dependencies initiated by potential risks, significantly affects maintenance-related decision-making. By considering the spatial variability of the equipment state and the stochastic evolution of failure dependencies under dynamic risk conditions, this paper presents a dynamic maintenance decision-making model based on an ontology-driven cloud dynamic Bayesian network (DBN) model and deep reinforcement learning (DRL). First, to explicitly model the complex dependencies and potential causes of equipment failure, a knowledge graph is established by constructing an ontology of the hierarchical structure, properties, and failure modes. Second, on the basis of the relationships among different entities, an ontology-driven cloud DBN model is developed to predict the degradation state space evolution under potential risk conditions. Finally, the probability outputs of the cloud DBN model under different degradation states, failure dependencies, and critical risks are integrated with various maintenance actions to develop a DRL approach for dynamic maintenance-related decision-making. The proposed methodology can address the time-varying degradation associated with failure dependencies and dynamic risks in the context of maintenance-related decision-making.