Online Data Preprocessing in the Adaptive Process Model Building Based on Plant Data

Accurate and efficient online measurements of process variables which give information about final product quality are necessary for process control and optimization. However, these process variables cannot often be measured by a sensor or the measurements are too expensive and/or not reliable enough and therefore are not used. The value of these difficult-to-measure variables is usually determined by laboratory analysis based on the samples taken from the process. This kind of measurement is performed periodically, with a long time delay in obtaining information, and it does not provide continuous monitoring of the final product quality and introduction of automatic control. To provide this, the estimation of the difficult-tomeasure process variables can be performed based on the process variables that are measured by sensors in the plant (so called easy-to-measure variables) and which correlate with difficult-to-measure variables [1]. For that it is necessary to have an appropriate mathematical model. In practice, the model is usually not available. Because industrial processes are generally quite complex to model, a rigorous theoretical modelling approach is often impractical, requiring a great amount of effort, or even impossible. Thus, obtaining the process model is based on the measured data [2]. In modern industrial plants there are hundreds of process variables which are measured and stored in the process database, so it is logical to use these data for process model building. However, the measured data taken from