The robust inferential estimation of polymer properties using stacked
neural networks is presented. Data for building non-linear models is r
e-sampled using bootstrap techniques to form a number of sets of train
ing and test data. For each data set, a neural network model is develo
ped which are then aggregated through principal component regression.
Model robustness is shown to be significantly improved as a direct con
sequence of using multiple neural network representations. Confidence
bands for the neural network model predictions also result directly fr
om the application of the bootstrap technique. The approach has been s
uccessfully applied to the building of software sensors for a batch po
lymerisation reactor.