|
Description:
The aim of this project is to develop a modular system for process monitoring and fault diagnosis. The goal is a software prototype for monitoring the process states and process equipment especially in chemical unit processes.
In the first part of the project control performance indices were developed and tested. The model predictive controller was implemented to control the quality of the product. Control performance calculations were implemented in the controller. A performance index that accounts for restrictions of controlled and manipulated variables was also developed. Indices were also developed for supervision of error signal and control moves. Using prediction error variance to evaluate control performance gave promising results. With prediction error variance, device faults were detected. The performance indiced applied to contol signals were able to detect aggressive control action. Indices developed for supervising process constraints were able to detect situations where the value of a controlled or manipulated variable was at its boundary.
The second part of the project consisted of a study of different fault detection and process monitoring methods applied in the chemical process industry and implementation of some applicable methods to monitor the dearomatization process. The methods were first tested with simulator data and later online with real process data from Fortum oil refinery. Artificial neural networks, fuzzy logic and statistical multivariable methods including principal component analysis and partial least squares were the selected methods to be used with a simulator.
Both self-organizing maps and projection onto latent structure models gave promising results. Principal component pre-processing clearly improved the SOMs’ ability to classify the process states. The drawback of the PLS model was its inability to identify the different abnormal states. In practice, the use of SOMs for process monitoring might be easier for a process operator, since the map with coloured neurons is very visual.
The results showed that variables have a significant effect on performance of the models. Monitored process variables should be selected based on both principal component analysis and process knowledge. Constructing computational variables that describe the monitored process phenomena is of great importance. The second part of the project also included an online fault diagnosis and model updating system for ethylene plant. The aim was to detect faults in analyzers and reconstruct missing measurement values.
The project was done in collaboration with Neste Engineering and the National Technology Agency of Finland, Tekes.
|
