Combined learning processes for injection moulding based on simulation and experimental data

Injection moulding enables the production of complex formed high-quality plastics parts in a single production step. To achieve a high and constant product quality, an appropriate process set-up with regards to product quality and process robustness is essential. A conventional process set-up requires expensive and time consuming experiments and know-how from the machine operator. One way to overcome this challenge is to make use of machine learning methods for process set-up. These methods can model the relationship between setting parameters and quality values and thus enable the identification of optimal working points. However, the training required for accurate modelling needs experimental data from extensive experiments for each process, too. Numerical simulation can predict quality values based on setting parameters without practical experiments. Whereas trends and the general dependencies between the parameters can be predicted with a satisfying accuracy, a certain discrepancy between the prediction of the simulation and the real process cannot be excluded. A combined approach using data from injection moulding simulations as well as experimental data appears promising to overcome the detriments of the solitary use of simulation for the training of machine learning algorithms. General dependencies could be gained from simulations without practical experiments and fine-tuning could be achieved by experimental trials with a minimal scope. In this paper, data obtained from a 2.5 D injection moulding simulation is compared with experimental data from a plate specimen and a complex formed injection moulded part. Therefore, central composed designs of experiments are used to identify differences in the effects and interdependencies of six setting parameters on quality values like part weight and dimensions. Furthermore, differences M the absolute values and the functionality of the effects are considered. On this basis, a combined machine learning concept using simulation and experimental data is presented.