Future manufacturing systems will have to become more intelligent to be able to guarantee a constantly high quality of products while simultaneously reducing labor-intensive quality-assurance tasks to address the shortage in workforce. In this work, we study the application of neural networks to the field of powder metallurgy and more specifically the production of green parts as part of a typical sintering process. More specifically, we explore the usage of neural-network-based predictions in closed-loop control. We train neural networks based on a series of produced workpieces, and use these networks in closed-loop production to predict quality characteristics like weight and dimensions of the workpiece in real-time. Based on these predictions an adaptive trajectory planner adjusts then trajectory key points and with this the final piston trajectories to bring and keep quality characteristics of workpieces within tolerance. We finally compare the control performance of this neural network-based approach with a pure sensor-based approach. Results indicate that both approaches are able to bring and keep quality characteristics within their tolerance limits, but that the neural network-based approach outperforms the sensor-based approach in the transient phase, whereas in steady state the neural network needed to be updated from time to time to reach the same high performance as the sensor-based approach. Since updating needs to be performed only from time to time, required expensive sensors can be shared among multiple machines and thus, costs can be reduced. At the same time the superior prediction performance of the neural-network-based approach in transient phases can be exploited to accelerate setting up times for new workpieces. Future work will target the automation of the recording of the training dataset, the exploration of further machine learning methods as well as the integration of additional sensor data to further improve predictions.
