Thesis for the Degree of Doctor of Technology


Year 2008
Mikko Vermasvuori

Methodology for utilising prior knowledge in constructing data-based process monitoring systems with an application to a dearomatisation process


Global competition is forcing the process industry to optimise the production processes. One key factor in optimisation is effective process state monitoring and fault detection. Another motivator to improve process monitoring systems are the substantial losses of revenue resulting from abnormal process conditions. It has been estimated that the petrochemical industry in the US alone loses 20 billion dollars per year because of unoptimal handling of abnormal process situations. Traditionally, the monitoring systems have been based on first principle models, constructed by specialists with process specific expertise. In contrast, the use of data-based modelling methods require less expertise and offers the possibilities to build and update the monitoring models in a short period of time, thus allowing more efficient development of monitoring systems.

The aims of this thesis are to augment data-driven modelling with existing process knowledge, to combine different data-based modelling methods, and to utilise calculated variables in modelling in order to improve the accuracy of fault detection and identification (FDI) and to provide all necessary diagnostic information for fault tolerant control. The suggested improvements are included in a methodology for setting up FDI systems. The methodology has been tested by building FDI systems for detecting faults in two online quality analysers in a simulated and in a real industrial dearomatisation process at the Naantali oil refinery (Neste Oil Oyj). In developing an FDI system, background information about the user requirements for the monitoring system is first acquired. The information is then analysed and suitable modelling methods are selected according to the guidelines given in the methodology. Second, the process data are prepared for the modelling methods and augmented with appropriate calculated variables. Next, the input variable sets are determined with the introduced method and the models are constructed. After the estimation accuracy of the models is validated, the values of the fault detection parameters are determined. Finally, the fault detection performance of the system is tested. The system was evaluated during a period of one month at the Naantali refinery in 2007. The monitoring system was able to detect all the introduced analyser faults and to provide the information needed for a fault tolerant control system, thus validating the methodology. The effects of a number of suggested improvements in data-based modelling are analysed by means of a comparison study.

Thesis electronical version can be downloaded from here

This info last modified 23 Sep 2017 by Jerri Kämpe-Hellenius