Due to the high labour costs in industrialised countries and the increasing demands on the quality of plastic moulded parts produced by the injection moulding process, companies are challenged to increase their production efficiency. Data-based approaches provide an opportunity for systematic and objective process set-up and control. Especially the process set-up can be a tedious, iterative and complex task which can consume a significant amount of time in the preparation of mass production. Artificial neural networks (ANN) can be used by evolutionary algorithms to search for suitable machine setting parameters for the desired part and process quality features. This often results in locally optimal solutions, which do not guarantee the best possible machine settings. Invertible neural networks (INN) could be a solution as with the help of these models, globally optimal solutions may be identified for definitive optimisation as the entire solution space for given quality parameters can be mapped in short time. In our approach, data for the implementation of the INN is collected in practical tests. The resulting weight and several geometric dimensions of the parts are determined as quality parameters. Based on these data, ANN are trained and utilized for the generation of synthetic data for the training of the INN. The creation of synthetic data is necessary because INN have millions of trainable parameters and therefore degrees of freedom which need to be determined using a bigger amount of training data in comparison to traditional feedforward ANN. In this paper, two different databases with a size of 10,000 samples each are being investigated. The two databases differ in the distributions of holding pressure and cooling times in order to investigate the influenceability of the prediction quality of the INN by training data. The sample settings for the injection velocity, packing pressure, mould temperature and melt temperature are uniformly distributed. Two different types of conventional and conditional invertible neural networks are implemented based on literature recommendations. Measured on validation data, the model quality (R² value) for the prediction of the part weight is 0.995 and 0.975 for the part dimensions. The machine setting parameters proposed by the INN can be analysed in regard to their impact. The holding pressure time is identified as the machine setting with the greatest influence on both part weight and part dimension. The predictions of the INN are validated in practical experiments with an R² value of 0.997 for the part weight. The weight of the parts produced with the proposed machine settings deviates on average by 0.24 % from the target.
Skip Nav Destination
Article navigation
28 May 2024
THE 2ND INTERNATIONAL SYMPOSIUM ON PLASTICS TECHNOLOGY
5–6 September 2022
Aachen, Germany
Research Article|
May 28 2024
Investigations on the applicability of invertible neural networks (INN) in the injection moulding process
Lukas Seifert;
Lukas Seifert
a)
1
Institute for Plastics Processing (IKV) in Industry and Craft at RWTH Aachen University
, Seffenter Weg 201, 52074 Aachen, Germany
a)Corresponding author: [email protected]
Search for other works by this author on:
Yannik Lockner;
Yannik Lockner
b)
1
Institute for Plastics Processing (IKV) in Industry and Craft at RWTH Aachen University
, Seffenter Weg 201, 52074 Aachen, Germany
Search for other works by this author on:
Christian Hopmann
Christian Hopmann
c)
1
Institute for Plastics Processing (IKV) in Industry and Craft at RWTH Aachen University
, Seffenter Weg 201, 52074 Aachen, Germany
Search for other works by this author on:
AIP Conf. Proc. 3012, 020001 (2024)
Citation
Lukas Seifert, Yannik Lockner, Christian Hopmann; Investigations on the applicability of invertible neural networks (INN) in the injection moulding process. AIP Conf. Proc. 28 May 2024; 3012 (1): 020001. https://doi.org/10.1063/5.0193494
Download citation file:
Pay-Per-View Access
$40.00
Sign In
You could not be signed in. Please check your credentials and make sure you have an active account and try again.
26
Views
Citing articles via
Inkjet- and flextrail-printing of silicon polymer-based inks for local passivating contacts
Zohreh Kiaee, Andreas Lösel, et al.
Effect of coupling agent type on the self-cleaning and anti-reflective behaviour of advance nanocoating for PV panels application
Taha Tareq Mohammed, Hadia Kadhim Judran, et al.
Design of a 100 MW solar power plant on wetland in Bangladesh
Apu Kowsar, Sumon Chandra Debnath, et al.
Related Content
Optimisation of warpage on thin shell part by using particle swarm optimisation (PSO)
AIP Conference Proceedings (September 2017)
Reaction coordinate flows for model reduction of molecular kinetics
J. Chem. Phys. (January 2024)
Optimization of shrinkage on thick plate plastic part by using particle swarm optimization (PSO)
AIP Conf. Proc. (November 2018)
Systematic control of collective variables learned from variational autoencoders
J. Chem. Phys. (September 2022)