Integration and comparison of neural networks and conventional modeling techniques

Artificial neural networks have re-emerged as a new, promising tool for engineering problems in recent years. The nonlinear approximation property and the learning capability of a neural network have been successfully applied to chemical process modeling and data analysis. However, from a statistical point of view, artificial neural networks are non-parametric models which can over-fit the data when the number of samples is limited. In the case of highly correlated, noisy inputs, the neural network approach can give poor generalization due to the fact that it minimizes the prediction bias but results in large variance. The study in this dissertation focuses on integrating neural networks with statistical methods for chemical process modeling and data analysis.; First, a neural net PLS (NNPLS) method is proposed which offers a framework for embedding artificial neural networks into PLS regression. The NNPLS method projects the original data down to a number of one-dimensional latent relations by PLS projection and then uses neural networks to learn the latent relations. The latent variables are proven to be orthogonal, which is useful for numerical diagnosis and further analysis. Second, the NNPLS method is extended to modeling dynamic systems by using a nonlinear ARX (auto-regressive with exogenous inputs) model and a nonlinear FIR (finite impulse response) model. Third, a variable selection approach using sensitivity analysis is proposed to select relevant variables in empirical modeling. It is shown that a nonlinear model with a smaller number of relevant variables can predict better than a model with more variables. The variable selection scheme is generalized to dynamic modeling where time-delays and time-lags of a dynamic model can be determined. Fourth, four learning algorithms for system identification using neural networks are compared. It is proven mathematically that pattern learning is a first order approximation of batch learning with respect to learning rates for feedforward networks, while for recurrent networks pattern learning is a zeroth order approximation of batch learning.; A number of industrial data bases are used to illustrate the effectiveness of the proposed methods in this dissertation. (Abstract shortened by UMI.)...