Development of An Adaptive Kalman Filter for Estimation in Chemical Plant.

Abstract

The application of the Kalman Filter to the on-line estimation of the state of a chemical process has so far met with a limited amount of success due to the inaccuracies and non-linearities of the mathematical models developed to describe the process being studied. To overcome these problems, two theoretical developments are proposed; the first being an adaptive form of the Kalman Filter which compensates for modelling errors by finding the mean and covariance of a number of "fictitious inputs" and the second a numerical technique for determining the state transition matrix of non-linear systems by the use of eigenvector theory. Following a series of off-line experiments, these modified forms of the Kalman Filter were applied to the on-line estimation of the state of a pilot plant scale double effect evaporator. The work involved can be conveniently divided into the following three sections. First of all, a seventh order mathematical model of the evaporator, together with suitable heat transfer correlations and compatible with the Kalman Filter, was derived and then tested by comparing the simulated responses with plant data. Secondly, two major software packages were developed: (i) The Hadios Executive Package, which was used for interactive data acquisition. (ii) The On-line Digital Filtering Package (OLDFP), a real-time operating system which controls the execution and data acquisition of filtering programs written in Fortran. Finally, a series of on-line filtering experiments were carried out, the results of which show that the proposed theoretical developments considerably improve the performance of the Kalman Filter by eliminating bias and divergence.