Zur Seitenansicht


Real-time CQA monitoring in bioprocesses: Generic model building workflows for chromatographic and dielectric spectroscopy data / submitted by Michal Sismis
AutorInnenSismis, Michal
Beurteiler / BeurteilerinBodenhofer, Ulrich
ErschienenLinz, 2018
Umfangix, 70 Blätter : Illustrationen
HochschulschriftUniversität Linz, Masterarbeit, 2018
URNurn:nbn:at:at-ubl:1-25703 Persistent Identifier (URN)
 Das Werk ist gemäß den "Hinweisen für BenützerInnen" verfügbar
Real-time CQA monitoring in bioprocesses: Generic model building workflows for chromatographic and dielectric spectroscopy data [6.69 mb]
Zusammenfassung (Englisch)

The aim of this master thesis was to develop a dashboard that simplifies the data analysis and provide a methodology for the systematic analysis of critical quality attributes of the biopharmaceuticals in real time. Two computational workflows were developed for these purposes. The first model was based on Phase I of process monitoring with the use of principal component analysis. The second one extends the usage of the first one and used partial least squares regression for outlier detection. Both models are based on a moving window approach. Moreover, the new methods were compared to traditional semi-supervised anomaly detection methods called the one-class support vector machine. For the testing purposes, three different chromatographic datasets were used. The simplified data analysis was achieved by the usage of interactive ipywidget sliders, which allow the user to select interesting regions within the data and adjust the size of the windows. Further, the condensation of the output from multi-channel analysis brought the intelligibility of the data analysis. The output of our both computational models were visualised by scatter plot and heatmap. Overall, five different experiments were conducted. The experiments were focused on the performance of our models, a capability of multi-channel data analysis, and investigation of detection functionalities of our models. Performance tests showed that our models can be considered as the trustworthy tools for systematic analysis of users data. In comparison to traditional machine learning methods, our model achieved identical performance for outlier detection, but the moving window approach passed the useful additional information about the data to the analyst. However, remaining experiments showed that our models are useful in the detection of anomalies in the analyzed data and are capable to pass this information to the user in the easily understandable way. Although there are available robust elements, which can be used in the model, the analysis of the dataset according to the general workflow of phase I in process monitoring can lead to the correct analysis of data. Moreover, it can also handle masking and swamping effects of outliers. Even with several drawbacks and possible further improvements to the model, we were able to develop the user-friendly interactive dashboard for the systematic analysis of data.

Das PDF-Dokument wurde 4 mal heruntergeladen.