Details
| Original language | English |
|---|---|
| Article number | 101822 |
| Number of pages | 20 |
| Journal | International Journal of Metalcasting |
| Early online date | 1 Apr 2025 |
| Publication status | E-pub ahead of print - 1 Apr 2025 |
Abstract
In this study, different machine learning algorithms were analysed to predict casting defects in a cold chamber magnesium high-pressure die casting process. Based on component-related process and quality data from 7982 casting cycles, models were trained using Support Vector Machine, Random Forest and AutoML algorithms from Auto-Sklearn. The aim was to predict the presence of defect classes ("cold flow", "shrinkage cavity", "blister", "soldering point", "scrap"). Random Forest models achieved the best prediction quality overall, especially for the defect class "soldering points", which is the least frequent detected defect. An analysis of training data amounts showed that the prediction quality only improves slightly beyond 1000 training cycles, except for the "soldering point" defect class, which showed further improvements with more training data. Furthermore, the scope of data in terms of measurement sources affected the prediction quality significantly. Random Forest prediction models that were trained exclusively with casting machine data generally provide a solid basis for predicting casting defects. The highest increase in prediction performance was achieved by adding die sensor data. Overall, the prediction quality of all models was always above the statistically expected values (Balanced Accuracy of 50%), and soldering points in particular were predicted with a Balanced Accuracy of more than 80%. It was found that block temperature sensors in the shot sleeve and force measurements (for measuring the cavity pressure via the ejector pin) had comparatively high correlations with all defect classes and were weighted highly by the Random Forest models for decision-making.
Keywords
- artificial intelligence, high-pressure die casting, industry 4.0, machine learning, quality prediction, random forest
ASJC Scopus subject areas
- Engineering(all)
- Mechanics of Materials
- Engineering(all)
- Industrial and Manufacturing Engineering
- Materials Science(all)
- Metals and Alloys
- Materials Science(all)
- Materials Chemistry
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: International Journal of Metalcasting, 01.04.2025.
Research output: Contribution to journal › Article › Research › peer review
}
TY - JOUR
T1 - Data-Driven Prediction of Casting Defects in Magnesium High-Pressure Die Casting Using Machine Learning
AU - Pachandrin, Slava
AU - Hoffmann, Norbert
AU - Dilger, Klaus
AU - Rokicki, Markus
AU - Niederée, Claudia
AU - Stürenburg, Lukas
AU - Noske, Hendrik
AU - Denkena, Berend
AU - Kallisch, Jonas
AU - Wunck, Christoph
N1 - Publisher Copyright: © The Author(s) 2025.
PY - 2025/4/1
Y1 - 2025/4/1
N2 - In this study, different machine learning algorithms were analysed to predict casting defects in a cold chamber magnesium high-pressure die casting process. Based on component-related process and quality data from 7982 casting cycles, models were trained using Support Vector Machine, Random Forest and AutoML algorithms from Auto-Sklearn. The aim was to predict the presence of defect classes ("cold flow", "shrinkage cavity", "blister", "soldering point", "scrap"). Random Forest models achieved the best prediction quality overall, especially for the defect class "soldering points", which is the least frequent detected defect. An analysis of training data amounts showed that the prediction quality only improves slightly beyond 1000 training cycles, except for the "soldering point" defect class, which showed further improvements with more training data. Furthermore, the scope of data in terms of measurement sources affected the prediction quality significantly. Random Forest prediction models that were trained exclusively with casting machine data generally provide a solid basis for predicting casting defects. The highest increase in prediction performance was achieved by adding die sensor data. Overall, the prediction quality of all models was always above the statistically expected values (Balanced Accuracy of 50%), and soldering points in particular were predicted with a Balanced Accuracy of more than 80%. It was found that block temperature sensors in the shot sleeve and force measurements (for measuring the cavity pressure via the ejector pin) had comparatively high correlations with all defect classes and were weighted highly by the Random Forest models for decision-making.
AB - In this study, different machine learning algorithms were analysed to predict casting defects in a cold chamber magnesium high-pressure die casting process. Based on component-related process and quality data from 7982 casting cycles, models were trained using Support Vector Machine, Random Forest and AutoML algorithms from Auto-Sklearn. The aim was to predict the presence of defect classes ("cold flow", "shrinkage cavity", "blister", "soldering point", "scrap"). Random Forest models achieved the best prediction quality overall, especially for the defect class "soldering points", which is the least frequent detected defect. An analysis of training data amounts showed that the prediction quality only improves slightly beyond 1000 training cycles, except for the "soldering point" defect class, which showed further improvements with more training data. Furthermore, the scope of data in terms of measurement sources affected the prediction quality significantly. Random Forest prediction models that were trained exclusively with casting machine data generally provide a solid basis for predicting casting defects. The highest increase in prediction performance was achieved by adding die sensor data. Overall, the prediction quality of all models was always above the statistically expected values (Balanced Accuracy of 50%), and soldering points in particular were predicted with a Balanced Accuracy of more than 80%. It was found that block temperature sensors in the shot sleeve and force measurements (for measuring the cavity pressure via the ejector pin) had comparatively high correlations with all defect classes and were weighted highly by the Random Forest models for decision-making.
KW - artificial intelligence
KW - high-pressure die casting
KW - industry 4.0
KW - machine learning
KW - quality prediction
KW - random forest
UR - http://www.scopus.com/inward/record.url?scp=105001936064&partnerID=8YFLogxK
U2 - 10.1007/s40962-025-01592-w
DO - 10.1007/s40962-025-01592-w
M3 - Article
AN - SCOPUS:105001936064
JO - International Journal of Metalcasting
JF - International Journal of Metalcasting
SN - 1939-5981
M1 - 101822
ER -