Simulation analytics for real-time repair decision-making on a production line using a digital twin

Easier access to real-time data gives the opportunity for simulation to support operational decision-making based on the current state of the actual system. We describe a problem from manufacturing where a maintenance team responds to downtime events by prioritising repair jobs to optimise the throughput of a production line. We propose the Aggregate Decision and State-space Multi-fidelity Metamodelling Approach (ADSMMA), a digital twin solution incorporating a discrete event simulation (DES) model and a neural network regression metamodel within a multi-fidelity simulation optimisation routine. Computational constraints on the number of inputs we can use for the metamodel restrict the detail with which we can describe the state of the system. We solve this by using the structure of the problem to guide an aggregation of the statespace to reduce the dimensionality of the system state variable while still maintaining a satisfactory level of predictability. Similarly, decisions for the full state-space are highdimensional, preventing a metamodel-based characterisation. This is greatly reduced by aggregating the decision space into a small set of policies. Finally, to train a metamodel for the reduced decision and state-space requires instantiating simulation model runs with the fully detailed state and decisions. We provide a means to do this and demonstrate our approach on a case study based on a real production line. Results suggest that ADSMMA makes the simulation optimisation feasible for digital twin. Where the metamodel is accurate, it can provide good results without running the high-fidelity DES model in the simulation optimisation step.