In their recent article, “Multi-agent deep reinforcement learning-based maintenance optimization for multi-dependent component systems”, MODAPTO partners Phuc Do and Alexandre Voisin, Benoit Iung from Université de Lorraine, CNRS, CRAN, alongside with Van-Thai Nguyen from Université de Lorraine, CNRS, CRAN and Waldomiro Alves Ferreira Neto from RANDOM, UFPE, Brazil, introduce a novel framework for optimizing maintenance strategies in complex systems.
According to the article, a multi-agent deep reinforcement learning-based CBM approach for a manufacturing system is developed considering both stochastic and economic dependence between components is proposed leveraging a state interactions model to accurately capture the stochastic dependencies between system components. The framework enables optimization of the maintenance actions. The research investigates the critical role of component dependencies in maintenance optimization and assesses the applicability of the proposed method across various contexts. Comparative analysis with traditional maintenance strategies highlights the potential benefits of the proposed approach.
As the authors mention, manufacturing systems consist of a set of interdependent components. However, addressing the dependence between these components remains a challenge in both maintenance modeling and the optimization process. In this paper, the authors propose a multi-agent deep reinforcement learning-based for maintenance planning optimization for a manufacturing system, taking into consideration both stochastic and economic dependencies between components. In this manner, they introduce a novel state interactions model, suggesting that the degradation state of one component may influence the degradation process of others. In addition, the maintenance planning optimization approach based on multi-agent deep reinforcement learning is developed considering both fully and partially observed states. The deployed multi-agent deep reinforcement algorithm, specifically Weighted QMIX, ensures scalability and efficient consideration of state interactions and economic dependencies between components. The feasibility and performance of the proposed maintenance approach are investigated through various numerical studies. When compared to traditional maintenance approaches, such as value iteration method, Dueling Double Deep Q Network, and Multi-Agent Deep Q Network, their proposed approach consistently demonstrates superior results.
The article was published in Expert Systems With Applications , a peer-reviewed international journal whose focus is on exchanging information relating to expert and intelligent systems applied in industry, government, and universities worldwide.