Development of a decision support system for profile extrusion.

One major challenge in profile extrusion is the prediction of shrinking and warpage, leading to high amounts of off-specification goods, especially during start-up and product change. The determination of optimal process parameters requires either long trials or highly experienced line operators. There are numerous contributing factors to shrinking and warpage both in the planning and operating phase. Not only the die design, but also specific temperatures in the extruder, die and cooling setup during production contribute to shrinking and warpage. Cross-domain collaborations in the Cluster of Excellence "Internet of Production" enable agile research concerning the enhancement of Industry 4.0 applications. Real-time data analysis with model reduction and machine learning is used to build an application-specific Digital Shadow, which can be used for the development of an app-based decision support system for line operators, helping them to quickly determine the optimal process parameters during operation. In a first step, a measurement system was developed, enabling the retrofit of existing analogue extrusion lines. This is not only useful for general quality management purposes, but also a necessary step for creating an interface between the extrusion plant and the industry 4.0 network containing the machine learning backend. For the collection of training data for the machine learning backend of the decision support system, a modular profile extrusion die with exchangeable end plates suitable for three different profiles was rheologically and thermally designed. In the three different profiles, asymmetric cooling behaviour leads to different degrees of warpage. A common data base is developed, comprised of live data from the extrusion line, archive data from previous extrusion trials and archive data from cooling simulations. A fourth data set is created using model order reduction methods. Collectively, this common data base lays groundwork for the development of an invertible neural network, which creates a current Digital Shadow selecting the appropriate data from the four data sets. In this paper, the suitability of data sampling and preprocessing methods for model order reduction are examined and requirements for the archive data are determined. [ABSTRACT FROM AUTHOR]