Unscented Kalman Filter. Application of the robust approach to polymerization processes

Abstract

The control of polymerization processes has central importance because operational conditions affect the processing and end-use properties of the product. The nonlinear controllers based upon rigorous models make use of the on-line state estimates obtained from the available measurements. For polymerization processes, the Unscented Kalman Filter has shown a rewarding performance for state estimation. Because the presence of outliers distorts the behaviour of the filter, Robust Statistics-based approaches have been proposed to reduce their detrimental effect on variable estimates. Until now, only Huber type M-estimators have been used as loss function of the estimation problem. In this work, the ability of other types of M-estimators to improve estimate robustness without introducing numerical problems is analysed. The performances of the M-estimators are compared for a copolymerization process within the framework of a filtering technique based on the Unscented Transformation, which uses a reformulation of the covariance of measurements errors.