Discontinuous Manufacturing Control Using Surrogate and RL Models

The patent creates a digital twin (virtual model) of the manufacturing line that replicates its behavior and parameters. This virtual copy allows for extensive parameter space exploration and control optimization without affecting the physical system, thereby maintaining control reliability while avoiding the complexity costs of exploring high-dimensional parameter spaces on the actual manufacturing line.

The patent performs control optimization and parameter exploration in advance using the digital twin before implementing changes on the actual manufacturing line. This preliminary action allows thorough investigation of parameter spaces virtually, ensuring reliable control strategies are developed beforehand without incurring the costs during actual production operations.

The patent uses a digital twin as a virtual copy of the manufacturing line to generate simulation data. This digital copy enables rapid data generation without the time constraints of physical experimentation, while the computational model is optimized to balance accuracy with computational efficiency, reducing the energy cost of simulations.

The patent employs transfer learning to adapt pre-trained models to specific manufacturing contexts by adjusting parameters rather than performing complete retraining. This approach significantly reduces computational costs while maintaining accurate predictions, as the model leverages existing knowledge from the digital twin and adapts it to specific scenarios with minimal additional computation.

The patent uses the digital twin to generate additional synthetic training data that complements actual production data. This virtual data source enables extensive training iterations without consuming physical materials or limiting the number of experiments, while maintaining model accuracy through the realistic behavior simulation of the digital twin.

The patent creates a unified training framework that simultaneously utilizes both actual production data and digital twin simulation data. This multi-functional approach allows the system to leverage the realism of actual data and the unlimited variability of simulation data together, improving model accuracy while enabling extensive training iterations without additional material consumption.

The patent uses the digital twin to pre-test and validate control strategies that push boundary conditions before implementing them on the physical manufacturing line. The digital twin acts as a cushioning layer that absorbs the risk of physically infeasible parameters, allowing extensive exploration of parameter ranges including extreme conditions without compromising the reliability or feasibility of actual production operations.

The digital twin serves as an intermediary between theoretical parameter exploration and physical implementation. It mediates by translating extreme or boundary parameters into validated control strategies that have been tested in the virtual environment, ensuring that only physically feasible and reliable parameters are transferred to the actual manufacturing line while maintaining broad parameter range coverage.