The inherent complexities of the extrusion process have made the development of both mechanistic and parametric models problematic. This contribution addresses the issues involved in developing a realistic model of an industrial reactive plasticating extruder to enable prediction of extrudate viscosity, which provides a good measure of product quality for the process. The complex nonlinearities associated with the process input-output mapping suggest that neural networks could be an appropriate modelling paradigm. However, the large number of parameters that had to be used caused problems during model identification, since only a limited data set was available. Resampling techniques were therefore used for model identification and validation, due to their efficient use of data and their ability to provide realistic inference of the true error rate associated with the identified models. The statistics obtained are utilised for network structure selection, outlier detection and the derivation of a distribution for model prediction errors. A final network model is presented with fixed confidence bounds, the weights of this network are analysed and an input-output mapping of the process is generated.
[1]
Elie Bienenstock,et al.
Neural Networks and the Bias/Variance Dilemma
,
1992,
Neural Computation.
[2]
R. Tibshirani,et al.
An introduction to the bootstrap
,
1993
.
[3]
Christopher M. Bishop,et al.
Neural networks for pattern recognition
,
1995
.
[4]
A. J. Morris,et al.
Artificial neural networks in process engineering
,
1991
.
[5]
R. Palmer,et al.
Introduction to the theory of neural computation
,
1994,
The advanced book program.
[6]
M. R. Mickey,et al.
Estimation of Error Rates in Discriminant Analysis
,
1968
.