Many variables are measured on-line in various processes, and this can produce data draining for the operators. One way to extract information from process measurements is to use multivariate methods in monitoring the process. This thesis presents an approach to constructing robust models from historical data, without having to conduct designed experiments. This is achieved by using data from at least one year to cover process variation and by validating the model with external data; objects are selected according to a criteria function.

Linear (PLS) and non-linear (ANN) models are compared in terms of their ability to monitor and predict. PLS models were best for monitoring, because they detect process deviations early; on the other hand, ANN models performed better in prediction, due to their ability to handle signal errors.

Bearing this in mind, a multivariate model was created and used on-line to monitor paperboard manufacturing, and proved to be a tool much appreciated by operators. A prestudy of how the application could be further improved by augmenting it with a knowledge-based system was also performed.

In addition, a study was done in which VIS and NIR spectra were recorded online and used in satisfactorily predicting product properties.

Various multivariate methods are briefly described in the thesis. The methods are multivariate data analysis, artificial neural networks, knowledge-based systems,and design of experiments. Various multivariate methods can be used to solve the same type of problems. Since different methods have different advantages, there are no conflicts between the methods. What is important is to choose the right method for a specific problem; often a combination of several methods can perform better than any single method.