Soft-sensing using recurrent neural networks

Focuses on the use of a neural network-based soft-sensing concept for process variables which cannot be measured online. An approach is proposed which performs particularly well in circumstances where the relationship between the measurable process variables and the variables to be estimated is difficult to establish or in cases where there are no sufficient data to construct a model. The method involves the modeling of an alternative relationship and the use of an a priori knowledge from which the unmeasured variable may be determined. To demonstrate this, a simulation model of a drying drum is considered, whose purpose is to increase the percentage of dry substance contained in pressed pulp. Among the process variables, the dry substance content of the pressed pulp at the inlet of the drum is assumed to be online immeasurable. A recurrent neural network is trained to predict the dry substance content of the pulp at the outlet of the drum, which is considered as a measurable output variable. The estimation of the unmeasured variable is carried out based on the prediction of the recurrent network and an a priori knowledge about the effect of the unmeasured variable on the output variable in relation to a measured input variable.