News, Blog‎ > ‎

PhD defense of Laszlo Dobos, 7th of June, 2016, 1pm

posted May 20, 2016, 3:04 PM by János Abonyi   [ updated May 20, 2016, 3:05 PM ]
Development of Experimental Design Techniques for Analyzing and Optimization of Operating Technologies

The aim of this thesis is to introduce theoretical basics of different approaches which can support further the production process development, based on the extracted knowledge from process data. As selection of time-frame with a certain operation is the starting point in a further process investigation, Dynamic Principal Component Analysis (DPCA) based time-series segmentation approach is introduced in this thesis first. This new solution is resulted by integrating DPCA tools into the classical univariate time-series segmentation methodologies. It helps us to detect changes in the linear relationship of process variables, what can be caused by faults or misbehaves. This step can be the first one in the model-based process development since it is possible to neglect the operation ranges, which can ruin the prediction capability of the model. In other point of view, we can highlight problematic operation regimes and focus on finding root causes of them. When fault-free, linear operation segments have been selected, further segregation segregation of data segments is needed to find data slices with high information content in terms of model parameter identification. As tools of Optimal Experiment Design (OED) are appropriate for measuring the information content of process data, the goal oriented integration of OED tools and classical timeseries segmentation can handle the problem. Fisher information matrix is one of the basic tools of OED. It contains the partial derivatives of model output respect to model parameters when considering a particular input data sequence. A new, Fisher information matrix based time-series segmentation methodology has been developed to evaluate the information content of an input data slice. By using this tool, it becomes possible to select potentially the most valuable and informative time-series segments. This leads to the reduction of number of industrial experiments and their costs. In the end of the thesis a novel, economic objective function-oriented framework is introduced for tuning model predictive controllers to be able to exploit all the control potentials and at the meantime considering the physical and chemical limits of process.