Robust learning algorithms for nonlinear filtering

A nonlinear adaptive filter for use in steel mill applications is described. This filter takes the form of a generalised regression network and is used to remove eccentricities from a steel mill force signal. Online adaptation is achieved by means of standard recursive parameter update algorithms suitable for linear regression type models. It is demonstrated that second order methods can lead to severe parameter biasing effects and a more serious effect known as bursting, whereby the parameter estimation becomes numerically unstable. This paper discusses the above effects form a theoretical and practical viewpoint, and considers the suitability of several learning algorithms for the eccentricity filtering application. This analysis leads to a numerically robust filtering structure, the efficacy of which is demonstrated by means of results from a real steel rolling mill