Computers and Chemical Engineering 28 (2004) 487–499 On-line adaptive parameter estimator tuning Michael Shagalov, Hector Budman Department of Chemical Engineering, University of Waterloo, Waterloo, Ont., Canada N2L 3G1 Received 14 January 2003; received in revised form 15 July 2003; accepted 19 August 2003 Abstract This paper discusses a systematic technique for tuning a parameter estimator. The estimator used in this work is a discrete version of a projection algorithm estimator. Although the development is explained for linear models, the method is equally applicable to any type of linear or non-linear model that is linear with respect to the estimated parameters. The estimation scheme includes a first-order filter in the error dynamics equation and individual parameter adaptation equations. A simple analytical formula was derived for on-line tuning of the filter gain based on the maximisation of the decay of the norm of the output and parameter errors. The tuning of the parameter adaptation gains is also done based on the maximisation of the decay of this norm. The measurement noise problem was addressed using a specific dead zone approach. The algorithm performance is compared to the conventional RLS algorithm. © 2003 Elsevier Ltd. All rights reserved. Keywords: Adaptive estimation; Tuning; Lyapunov stability 1. Introduction Chemical systems are generally non-linear due to nonlin- earity of reaction rate expressions or non-linear dependency of physical parameters with respect to temperature, pressure or flow conditions. Often, these non-linear models cannot be easily generated using first principle equations and con- sequently, empirical non-linear models are used for estima- tion, model based filtering and control. New developments in neural networks (NNs) models based on wavelets or ra- dial basis functions have generated renew interest in the area of adaptive estimation and control of non-linear systems based on these novel empirical non-linear models (Sanner & Slotine, 1992a, 1992b; Kavchak & Budman, 1999). For example, a model based on non-linear basis functions is formulated as follows: y k = n i=1 a i f 1 (y k-i ) + m j=1 b j f 2 (u k-j ) (1.1) In order to represent the non-linear behaviour of the system different types of basis functions may be used for Corresponding author. Tel.: +1-519-888-4601; fax: +1-519-746-4979. E-mail address: hbudman@uwaterloo.ca (H. Budman). f 1 (y k-i ) andf 2 (u k-i ) such as Gaussians or Wavelet func- tions. In Eq. (1.1), y k -i and u k-j denote past output and input measurements, respectively. This superposition structure is generally referred to as a NN (Sanner & Slotine, 1992a, 1992b). The coefficients a i and b j are identified from experimental data. These coeffi- cients are often learned on-line because processes are either time-varying or non-linear so they will require a large num- ber of experiments at different operating points for complete off-line identification. When linear dynamics is considered, a discrete dynamic equation can be expressed as follows: y k = n i=1 a i y k-i + m j=1 b j u k-j (1.2) Thus, it is evident from Eqs. (1.1) and (1.2) that the pa- rameter estimation problem is similar for both the linear and non-linear models given by Eqs. (1.1) and (1.2), since both methods are linear with respect to the parameter set [a 1 ,a 2 ,...,a n ,b 1 ,b 2 ,...,b m ] to be learned. Eq. (1.2) may be seen as a special case of Eq. (1.1) when f(x) = x. This fact is important for the present study since the methodol- ogy in this work has been illustrated for the simpler linear case but is equally applicable to the non-linear models given by Eq. (1.1). 0098-1354/$ – see front matter © 2003 Elsevier Ltd. All rights reserved. doi:10.1016/j.compchemeng.2003.08.003