Building robust neural networks for industrial process applications

An alternative approach to the generation of realistic confidence bounds is proposed which considers two aspects of the data which are known to influence the resultant prediction error. These are the ability of the network to predict the output and the impact of the density of the training data on the resultant bounds. Neural network predictions encompassed by confidence bounds provide a more robust and believable representation of the forecast, and provide industries with greater confidence when using such models as software sensors and for monitoring and control. The power of the technique lies in the fact that it is not necessary to identify an underlying distribution which is difficult to satisfy and that it is equally applicable to all forms of network topologies and non-linear modelling procedures. The two techniques are applied to a pilot plant batch methyl methacrylate polymerisation reactor