Nonlinear Dimension Reduction Based Neural Modeling for Nonlinear Complex DPS

Abstract

A nonlinear principal component analysis (NL-PCA) based neural modeling approach is presented for a lower-order or more accurate solution for nonlinear distributed parameter systems (DPS). One NL-PCA network is trained for the nonlinear dimension reduction and the nonlinear time/space reconstruction. The other neural model is to learn the system dynamics with a linear/nonlinear separated model structure. With the powerful capability of dimension reduction and the intelligent learning, this approach can model the nonlinear complex DPS with much more complexity. The simulation on the catalytic rod and the experiment on the snap curing oven will demonstrate the effectiveness of the presented method.