Monitoring and control of laser-based additive manufacturing using convolutional neural nets and reinforcement learning

Directed Energy Deposition (DED) processes using industry robots enable the production of large structures through additive manufacturing technology. One major challenge hindering the industrialization of DED processes is online process monitoring and control. Melt pool characteristics are one of the key identifiers regarding the stability of DED processes. Typically, these characteristics are monitored visually by camera systems using standard image processing toolboxes. To improve the robustness, the image processing of melt pools has been replaced by a convolutional neural net. Furthermore, a control strategy, focusing on laser power and processing speed, has been introduced through a Reinforcement Learning Framework. Wherein, a second neural net, predicting melt pool characteristics, has been used as the simulation environment. Based on this, the reward metric for the reinforcement system has been determined by comparing target melt pool characteristics with the predicted ones.