Year 2004
Marjut Kinnunen

Integration of first-principles process models into data-based monitoring

Recently, process monitoring has been under active research. Emerging trends include combining different methods and also combining monitoring with control into fault tolerant control. The aim of this thesis is to study different ways of integrating deterministic models into a statistical multivariate monitoring method and to develop a method for evaluating estimate realiability that could later be used for control purposes.

The literature survey part of this study contains a short summary of the most popular multivariate methods and focuses on various ways of combining first-principles models with empirical identification methods found in the literature. Also, aspects of data reliability and reported reliability indexes are covered.

In the experimental part of the study, integration of a non-linear differential equation model, balance equations and other derived knowledge-based variables into a partial least squares regression model in an industrial monitoring application was studied. Process dynamics and nonlinearity were described by the dynamic model and derived variables, which produced input to the PLS model. The dependency between PLS model inputs and outputs could then be modelled as linear. Balance equations in turn were used to separate the variance due to several different operation schemes in the application process from the PLS model inputs in order to make the monitoring more sensitive to the relevant variance around different operating points. Also, based on examples reported in the literature, an estimate reliability index calculated from input residuals weighed by model weights was developed. The index was used to eliminate unreliable estimate effect in monitoring where it was based on prediction error.

It was shown that a hybrid model of this type is more robust in, for example, production volume changes than purely statistical model and that the nuisance variance due to several products was successfully removed from the input data. The hybrid monitoring application performed better than the purely statistical application: 97 % instead of 93 % of faulty analyzer measurements were detected and the hybrid system produced 52 % less false alarms. With the connection from the reliability index, the amount of false alarms further decreased to one tenth.

This info last modified 27 Apr 2024 - contact webmaster