Digital Twin-Based Performance Optimization of Thermal Systems in Smart Manufacturing

Abstract

One of the key issues associated with the intelligent manufacturing is the thermal management. The temperature homogeneity and efficiency define the quality of goods to be produced as well as the production costs. The authors suggest a novel intelligent hybrid digital twin architecture which allows for real-time optimization of the industrial thermal processes and integrates the physics-informed reduced-order model of thermal processes and the deep reinforcement learning algorithm for the closed-loop control of the system. There are two main problems that can be solved using this architecture; the first problem is the computational time required for the high-fidelity simulations while the second problem is the insufficiency of the necessary data required for the adaptive decision-making process in the physical system. Spatiotemporally reduced-order modeling is applied for the fast estimation of the temperature field on the basis of the data collected by the sensors while the deep Q-Network is controlling the temperature distribution according to several objectives.