Local linear neural networks based on principal component analysis

Abstract

A new method for the estimation of process parameters based on the principal component analysis is developed. The estimator yields optimal estimation results in the case of errors in variables (EIV) problems which are characterized by corrupted measurements of input and output signals. As the residual generation in fault detection methods often feature EIV characteristics, the estimator can be used to identify linear models for residual calculation. To overcome the limitations on linear models, the developed estimator is integrated into the LOLIMOT approach which is able to identify nonlinear processes. The estimator is used as an alternative to the standard Least Squares estimator to identify the parameters of the local linear models. Comparative results show the better suitability of the developed estimator for the residual generation in EIV-setups.