Methodologies for structural analysis from mass-data in scale plants

M.Eng. David Arengas

March 2014 - March 2019

Colfuturo - German Academic Exchange Service (DAAD) - Department of Measurement and Control

Design and implementation of experiments are required to collect data which is informative for system identification.

Performing experiments can be undesirable or even prohibited e.g. in running production processes since they can deviate a system from a pursued operating point. The production can be negatively impacted e.g. regarding product quality or safe operation. Such situations are frequently found in continuously operated plants e.g. in the chemical industry. Process data is usually logged for large periods of time and it can be used for system identification. However, signals found in such processes are predominantly stationary. The model accuracy can degrade if such as data sets are considered as a whole for system identification since the data is poorly informative. Therefore, these data shall first be “mined” for suitable data sequences to avoid model accuracy degradation. More accurate models can be obtained if just the informatively “richest” sequences of the original data set are used for parameter estimation.

Data mining for identification can be performed based on a model structure. The selected model class shall be flexible enough because few knowledge about the process can be available.

 Scope of the project:

• Data to identify multivariable models

• Data from open-loop as well as closed-loop operated plants

The search method to be developed will be tested and demonstrated on data from benchmark processes as well as data collected in the model factory µPlant of the Department of Measurement and Control.