Data-Driven Early Quality Prediction in High Pressure Die Casting

High pressure die casting (HPDC) is a complex manufacturing process often followed by several additional processing steps. Detecting faulty cast parts immediately after extraction from the die allows for the elimination of subsequent, resource-intensive processing steps for parts that will ultimately be rejected. Currently, casting error detection is typically based on tolerance thresholds for manufacturing process parameters. If a single parameter falls outside the specified range, the cast part may be scrapped before proceeding to subsequent production steps. This tolerance check is often integrated directly into the HPDC machine monitoring systems. In order to improve this method, we present a data-driven approach for the early identification of casting errors based solely on HPDC machine data available immediately after the casting. Our model is trained on process data from an industrial foundry, obtained during regular production. We show that our method surpasses the classical tolerance-based rejection method based on additional tests, which were performed independent of the training data and with manual quality inspection of each casting part to compare the actual quality with the predictions. The economic impact of the suggested improvements over the state of the art is briefly discussed as well.