Improvement of prediction performance for data-driven virtual sensors

Virtual sensors (VS) allow measurement of process parameters where direct measurement is too expensive or even not possible. For the virtual sensors which build their internal process model after the data-driven method, e.g. by use of an artificial neural network (ANN), there is a problem of the evaluation of the prediction performance. The up to date solutions solve this problem only partially and only for few ANN types, require huge development effort and are inapplicable for the real time operation. A new approach for the improvement of the VS prediction performance based on the statistical process control (SPC) methods is suggested in this article. It is valid for a wide class of the ANN and reduces the development effort severely. The simulation of this approach using the real process data has delivered promising results.