Data Mining of Process Data in Multivariable Systems

Sammanfattning: Performing system identification experiments in order to model control plantsin industry processes can be costly and time consuming. Therefore, with increasinglymore computational power available and abundant access to loggedhistorical data from plants, data mining algorithms have become more appealing.This thesis focuses on evaluating a data mining algorithm for multivariate processwhere the mined data can potentially be used for system identification.The first part of the thesis explores the effect many of the necessary user chosenparameters have on the algorithmic performance. In order to do this, a GUIdesigned with assisting in parameter selection is developed. The second partof the thesis evaluates the proposed algorithm’s performance by modelling asimulated process based on intervals found by the algorithm.The results show that the algorithm is particularly sensitive to the choice ofcut-off frequencies in the bandpass filter, threshold of the reciprocal conditionnumber and the Laguerre filter order. It is also shown that with the GUI itis possible to select parameters such that the algorithm performs satisfactoryand mines data relevant for system identification. Finally, the results show thatit’s possible to use the mined data in order to model a simulated process usingsystem identification techniques with good accuracy.