Bitte benutzen Sie diese Kennung, um auf die Ressource zu verweisen: http://dx.doi.org/10.18419/opus-9977
Langanzeige der Metadaten
DC ElementWertSprache
dc.contributor.authorScheller, Stefan-
dc.date.accessioned2018-08-23T09:38:26Z-
dc.date.available2018-08-23T09:38:26Z-
dc.date.issued2018de
dc.identifier.other510551394-
dc.identifier.urihttp://nbn-resolving.de/urn:nbn:de:bsz:93-opus-ds-99941de
dc.identifier.urihttp://elib.uni-stuttgart.de/handle/11682/9994-
dc.identifier.urihttp://dx.doi.org/10.18419/opus-9977-
dc.description.abstractSimulations of physical processes are of great importance in many areas of research. Typically, time-dependent volume data with high resolution are generated in this context. Simulations are often repeated multiple times with different setup parameters, creating an ensemble data set containing several instances of volumetric data. The high-dimensional nature of ensemble data prevents the application of direct visualization methods and motivates automated techniques that support the visualization of the simulation results. This thesis deals with the analysis of spatiotemporal velocity data of a liquid flowing through a channel containing a cylinder. We present a method for the extraction of characteristic representations based on artificial neural networks. For this purpose two different types of representations are studied: time steps which contain the velocity vector field of the liquid flow at a specific point in time and isolines which mark velocity vectors of a certain length. In addition, results from different simulation runs are compared pairwise and their similarity is evaluated with a distance metric. Subsets of the simulation data, in form of the two representations, and the calculated distance values serve as input and target output for the supervised learning of the neural networks. For learning the distance metric, we present a convolutional neural network whose architecture is adapted to the significant size of the input data, the use of different amounts of representations and the symmetry of the metric. The trained networks are used to predict the distances between simulations of a separate evaluation data set. The resulting prediction accuracy serves as measure for the information content of the representations that were used for the training. In addition to the technique of extracting characteristic representations, we present methods for visualizing time steps and isolines over the entire time series of a simulation. The effectiveness of the extraction method is discussed in a comparison of visualizations resulting from those representations, which have achieved the highest and lowest prediction accuracy.en
dc.language.isoende
dc.rightsinfo:eu-repo/semantics/openAccessde
dc.subject.ddc004de
dc.titleAnalysis of spatiotemporal ensemble data using machine learningen
dc.typemasterThesisde
ubs.fakultaetInformatik, Elektrotechnik und Informationstechnikde
ubs.institutInstitut für Visualisierung und Interaktive Systemede
ubs.publikation.seiten74de
ubs.publikation.typAbschlussarbeit (Master)de
Enthalten in den Sammlungen:05 Fakultät Informatik, Elektrotechnik und Informationstechnik

Dateien zu dieser Ressource:
Datei Beschreibung GrößeFormat 
AnalysisOfSpatiotemporalEnsembleDataUsingMachineLearning.pdf7,95 MBAdobe PDFÖffnen/Anzeigen


Alle Ressourcen in diesem Repositorium sind urheberrechtlich geschützt.