Applying Statistical Methods for Detection of Porosity in Direct Metal Laser Sintering Parts
Raitanen, Niko (2018)
2018
Materiaalitekniikka
Teknisten tieteiden tiedekunta - Faculty of Engineering Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2018-02-07
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201801231139
https://urn.fi/URN:NBN:fi:tty-201801231139
Tiivistelmä
Additive manufacturing (AM) has established itself as a technology able to provide efficiency not only in prototyping, but also in serial production. However, the development of in situ real-time monitoring and closed-loop system is ongoing. This thesis aims to evaluate applicability of data-driven modeling with process monitoring data. The first research question was to define the ability of commercial EOSTATE MeltPool monitoring system to distinguish faulty process conditions from standard process of EOS NickelAlloy IN718 material. Then, the possibility to build a predictive regression model for porosity detection was evaluated.
The literature review concentrates on defects and porosity forming mechanisms in powder bed fusion process. Experimental part consists of processing parts simultaneously collecting monitoring data and subsequently analyzing the parts in laboratory. Parts were produced containing layers of both standard process and provoked process simulating issues with shield gas flow. Another set of parts was built with variation in other important process parameters to evaluate correlation between quality degradation in samples and monitoring signal. The quality of the parts was quantified by measuring defect fractions of the samples in laboratory and deviations were statistically analyzed. State-of-the-art statistical learning algorithms were used to construct a predictive model for porosity detection.
The differences in signal between standard and provoked processes were statistically tested to conclude that EOSTATE MeltPool algorithms can differentiate ill-fated shield gas flow from normal process conditions. Correlations between measured monitoring data and defect fraction of the produced parts were evaluated. Non-linear statistical learning algorithms, like support vector machine with radial kernel extension and random forest, reached a statistically significant accuracy in fitting a regression model to predict defect fraction of parts from process monitoring signal. The predictive models were tested with another set of built parts to confirm that the monitoring signal is reproducible and repeatable. As a result, this thesis presents a concept for quality assurance of metal powder bed fusion process based on statistical learning.
The literature review concentrates on defects and porosity forming mechanisms in powder bed fusion process. Experimental part consists of processing parts simultaneously collecting monitoring data and subsequently analyzing the parts in laboratory. Parts were produced containing layers of both standard process and provoked process simulating issues with shield gas flow. Another set of parts was built with variation in other important process parameters to evaluate correlation between quality degradation in samples and monitoring signal. The quality of the parts was quantified by measuring defect fractions of the samples in laboratory and deviations were statistically analyzed. State-of-the-art statistical learning algorithms were used to construct a predictive model for porosity detection.
The differences in signal between standard and provoked processes were statistically tested to conclude that EOSTATE MeltPool algorithms can differentiate ill-fated shield gas flow from normal process conditions. Correlations between measured monitoring data and defect fraction of the produced parts were evaluated. Non-linear statistical learning algorithms, like support vector machine with radial kernel extension and random forest, reached a statistically significant accuracy in fitting a regression model to predict defect fraction of parts from process monitoring signal. The predictive models were tested with another set of built parts to confirm that the monitoring signal is reproducible and repeatable. As a result, this thesis presents a concept for quality assurance of metal powder bed fusion process based on statistical learning.