The dynamic modelling of ill-defined processing operations using connectionist networks
This paper porposes a new methodology to model ill-defined processing operations using neural nets (NNs). A process with many variables (large dimensionality) cannot be modelled adequately if limited process data are available. This problem of multidimensionality is addressed and an approach suggested to reduce the dimensionality using NNs. An NN is trained on process data for the global variable space, whereafter the first-order partial derivatives of the process are estimated with the NN and used to perform perturbation analyses. As a result, the variable space can be subdivided into subspaces with reduced dimensionality. The final product of this modelling methodology is a combined model of phenomenological expressions and NNs. The model can be incorporated successfully in a dynamic simulator of the process. A new approach to conduct modelling on the basis of continuous data collected directly from an industrial processing unit is also proposed. The modelling methodology is applied to a typical processing operation, i.e. the carbon-in-leach (CIL) process. © 1993.