Multimode approach using Reinforcement Learning and Digital Twin for operating mode management

Abstract Managing operating modes in a multimode production system represents a complex challenge that necessitates both meticulous and reactive planning. The challenge resides in coordinating and optimizing the different modes to address variations in demand while ensuring optimum makespan. A multimode system integrates multiple operating modes to accommodate any disturbances that may affect the system. This paper addresses the issue of selecting the appropriate mode to activate in response to the occurrence of a failure. By using Reinforcement Learning (RL) and Digital Twin (DT), the RL agent uses a state space (S t ) provided by a Digital Twin, to target its action (A) which consists of making a decision about which mode should be activated and which modes should be deactivated. The combination of the RL agent with the multimode system via the Digital Twin enables real-time adaptation to a dynamic environment, with the possibility of virtually testing the decisions made by the RL agent before their actual implementation, and consequently enhancing the performance of complex industrial systems. An innovative multimode scheduling approach will be targeted for the discrete case.