Aiming for the minimum number of training data for MIMO control in injection molding
Sprache des Vortragstitels:
Englisch
Original Tagungtitel:
European regional meeting of the Polymer Processing Society, PPS-2024
Sprache des Tagungstitel:
Englisch
Original Kurzfassung:
Injection molding is a well-known automatable technology annually converting more than 60 million tons of plastics to parts. However, IM is still suffering from scrap rates up to 1% or more. Being able to monitor and control the plastics? optimal transient process conditions in the mold cavity would support a constant, disturbance-insensitive, and machine-independent part quality. For this, a digital process twin of the manufacturing cell is aimed, which needs a comprehensive process model, transient process-states monitored by trustful soft and hard sensors in mold, machine, auxiliaries, and environment, OPC UA communication in between, and closed-loop (semi-)autonomic control algorithms. One approach is to pretrain these process-sensor-performance models using numerical CFD simulation studies.
Supervised machine learning typically profits from a multitude of experiments and a low number of corresponding outputs. This paper asks if a small set of 26 random design points, simulated in Cadmould v11, is sufficient for training regression models to predict the part qualities within the whole recommended processing window well. Melt temperature, injection rate, packing time and pressure, A-side and B-side coolant temperatures, and cooling time, were iterated. Mass, mean shrinkage, 2 planar lengths, 2 curved widths, max. deformation, and max. warpage of a curved-surface ABS part were determined as outputs, respectively.
Employing Orange 3.37 ? a visual programming data mining Python application from University of Ljubljana ? models using multi-layer perceptron (MLP) backpropagation ANN, Support vector machines SVM, AdaBoost, Linear Lasso, Random Forest, and Logistic, Linear, and Partial Least Squares Regression algorithms were trained, benchmarked, and tested on a set of 79 unknown design points.
In conclusion, the predictive power of several algorithms was high but still insufficient.
Sprache der Kurzfassung:
Englisch
Englischer Vortragstitel:
Aiming for the minimum number of training data for MIMO control in injection molding
Forschungs-