TY - JOUR TI - Maintenance modeling for degrading systems with individually repairable components using optimization and reinforcement learning DO - https://doi.org/doi:10.7282/t3-kzkc-3194 PY - 2021 AB - There are many different industrial and manufacturing multi-components systems, where each component experiences multiple competing failure processes, such as degradation and environmental shock. In this research there are two competing failure processes for each component, namely soft failure due to the degradation and hard failure due to the random shocks coming to the system. Moreover, each incoming shock results in damage on all the degradation paths of all the components. For some multi-component systems, components can be repaired/replaced individually within the system. For such multi-component systems with individually repairable components and dependent failure processes, it is not economical to replace the whole system if it fails, while in most of the previous studies, the systems are either considered to have independent failure processes or packaged and sealed together, and the whole system was replaced with a new one when any component fails. For systems functioning for a very long time, and each component is repairable within the system, individual components have been replaced several times. Therefore, the starting time or age of all the components within the system are not the same. In this research work, the conditional reliability analysis of such systems is studied considering the initial age of each component, at the beginning of the inspection intervals, as a random variable. For systems, whose costs due to failure are high, it is prudent to avoid the event of failure, i.e., the components should be repaired or replaced before the failure happens. Condition-based maintenance models recommend a policy to initiate repair or replacement before the failure occurs by detecting the system degradation status at each inspection time interval. Different types of condition-based maintenance models, including both static and dynamic models, are formulated and optimized to find the best maintenance policy. In some of the proposed models, determination of the optimal maintenance thresholds, such as on-condition and opportunistic thresholds for each component, along with optimal inspection time for the whole system, work to prevent failures and minimize cost. For multi-component systems with repairable components, it is also beneficial to have a dynamic maintenance plan based on the current degradation level of all the components in the system. In this study, different dynamic condition-based maintenance models are proposed using optimization and reinforcement learning methods. Moreover, different types of maintenance actions and the uncertainty of the maintenance implementation are also considered in the formulation of the maintenance models. Using a reinforcement learning approach provides a more time-efficient and cost-effective method compared to the traditional maintenance optimization solutions, and it can also provide a dynamic maintenance policy for each specific degradation state of the system. This can be more useful and beneficial compared to the fixed or stationary maintenance plans. The maintenance problems are formulated as a Markov decision process and are solved by using a Q-learning algorithm and deep Q-learning. Overall, the goal of the proposed research is to provide practical and effective maintenance models for a multi-component system with individually repairable components to avoid the failure and high downtime cost, and to minimize the cost KW - Materials -- Fatigue KW - Industrial and Systems Engineering LA - English ER -