An efficient data-driven method for the digital twin prediction update synchronization problem

Determining how to manage the trade-off between aligning the digital twin with its physical system and keeping the old prediction is crucial in dynamic environments. The digital twin prediction update synchronization problem determines whether to update the performance prediction from the digital twin based on the observed state of the physical system. Although existing methods provide solutions, they are limited to simple systems and constrained by the dependency on digital twin–generated sample paths. This study presents a data-driven method using actual data to solve the problem for complex systems. The method estimates the performance measure under consideration by observing limited information of the system to determine the optimal synchronization policy that balances synchronization costs and prediction bias. An optimal state-dependent policy and a periodic state-dependent policy are determined to indicate when to update the prediction. An unreliable production line is illustrated as a specific system to demonstrate the problem and its solutions. Our experimental results show that the data-driven method significantly improves computational efficiency while providing comparable solutions to the standard simulation method. Furthermore, focusing on limited information of the system, such as the state of the bottleneck machine in a production line, is efficient for solving the problem without requiring longer sample paths.