A reinforcement learning decision model for online process parameters optimization from offline data in injection molding

Injection molding is widely used owing to its ability to form high precision products. Good dimensional accuracy control depends on appropriate process parameters settings. However, existing optimization methods fail in producing ultra-high precision products due to their narrow process windows. In order to address the problem, an online decision system which consists of a novel reinforcement learning framework and a self-prediction artificial neural network model is developed. This decision system utilizes the knowledge learned from offline data to dynamically optimize the process of ultra-high precision products. Process optimization of an optical lens is dedicated to validating the proposed system. The experimental results show that the proposed system has excellent convergence performance in producing lens with deviation not exceeding ± 5μm. Comparison with the static optimization method prove that the decision model is more robust and effective in online production environment. And it achieves superior results in continuous production with the process capability index of 1.720 compared to 0.315 in fuzzy inference system. There is great potential for utilizing the proposed data-driven decision system in similar manufacturing process.