The contents of this thesis can be divided into two main parts. The first one is the development of an identification methodology for the modelling of complex industrial processes. The second one is the application of this methodology to the curl and twist problem.
The main purpose behind the proposed methodology is to provide a schematic planning, together with some suggested tools, when confronted with the challenge of building a complex model of an industrial process. Particular attention has been placed to outlier detection and data analysis when building a model from old, or historical, process data.
Another aspect carefully handled in the proposed methodology is the identifiability analysis. In fact, it is rather common in process modelling that the model structure turns out to be weakly identifiable. Thus, the problem of variable selection has been studied in this work, and a new algorithm for variable selection based on regularization has been proposed and compared with some of the classical methods, yielding promising results.
The second part of the thesis is about the development of a curl predictor. Curl is the tendency of paper of assuming a curved shape and is observed mainly during humidity changes. Curl in paper and in paperboard is a long-standing problem because it may seriously affect the processing of the paper. Unfortunately, curl cannot be measured online, but only in the laboratory after that an entire tambour has been produced. The main goal of this project is then to develop a model for curl and twist, and eventually to implement it as an on-line predictor to be used by the operators and process engineers as a tool for decision/control.
The approach we used to tackle this problem is based on grey-box modelling. The reasons for such an approach is that the physical process is very complex and nonlinear. The influence of some inputs is not entirely understood, and besides it depends on a number of unknown parameters and unmodelled/unmesurable disturbances.
Simulations on real data show a good agreement with the measurement, particularly for MD and CD curl, and hence we believe that the model has an usable accuracy for being implemented as an on-line predictor.