The viscosity of polymer melts is an important parameter for many applications. Several possibilities can be used to determine the viscosity parameters, but most of them are suitable for laboratory use only. A fast information about the viscosity parameters is necessary to ensure the quality in high volume productions of polymers, blends and compounds. In this case all of the laboratory methods are not qualified due to their feedback time. To address this problem a real-time determination of the viscosity parameters can be established using a soft sensor. Our work is focused on the development of a robust soft sensor based on models designed with artificial neural networks. In order to simulate slight material variations e.g. caused by batch changes we have manipulated the material properties by adding small amounts of material with a slightly altered molecular structure. This offers the possibility to observe the influence of batch changes on the modeling which reflects the quality of the soft sensor. Using representative data for modeling the prediction quality in slightly changing systems can be improved. Moreover, also the quality and usability of the soft sensors can be enhanced.